Initial commit: FunASR Speech Recognition Toolkit

Add complete FunASR codebase including models, runtime, and documentation.

funasr/datasets/audio_datasets/datasets.py

@@ -0,0 +1,330 @@
+import torch
+import random
+
+
+from funasr.register import tables
+from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+
+
+@tables.register("dataset_classes", "AudioDataset")
+class AudioDataset(torch.utils.data.Dataset):
+    """
+    AudioDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        is_training: bool = True,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs,
+    ):
+        """Initialize AudioDataset.
+        
+            Args:
+                path: TODO.
+                index_ds: TODO.
+                frontend: Audio frontend for feature extraction.
+                tokenizer: Tokenizer instance for text encoding/decoding.
+                is_training: Boolean flag for training.
+                int_pad_value: TODO.
+                float_pad_value: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+
+        self.preprocessor_speech = None
+        self.preprocessor_text = None
+
+        if is_training:
+            preprocessor_speech = kwargs.get("preprocessor_speech", None)
+            if preprocessor_speech:
+                preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+                preprocessor_speech = preprocessor_speech_class(
+                    **kwargs.get("preprocessor_speech_conf")
+                )
+            self.preprocessor_speech = preprocessor_speech
+            preprocessor_text = kwargs.get("preprocessor_text", None)
+            if preprocessor_text:
+                preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+                preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf"))
+            self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.int_pad_value = int_pad_value
+        self.float_pad_value = float_pad_value
+
+    def get_source_len(self, index):
+        """Get source len.
+        
+            Args:
+                index: TODO.
+            """
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        """Get target len.
+        
+            Args:
+                index: TODO.
+            """
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        """Internal: len  ."""
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        """Internal: getitem  .
+        
+            Args:
+                index: TODO.
+            """
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+
+        if self.tokenizer:
+            ids = self.tokenizer.encode(target)
+            text = torch.tensor(ids, dtype=torch.int64)
+        else:
+            ids = target
+            text = ids
+        ids_lengths = len(ids)
+        text_lengths = torch.tensor([ids_lengths], dtype=torch.int32)
+
+        return {
+            "speech": speech[0, :, :],
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+        }
+
+    def collator(self, samples: list = None):
+        """Collator.
+        
+            Args:
+                samples: TODO.
+            """
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs
+
+
+@tables.register("dataset_classes", "AudioDatasetHotword")
+class AudioDatasetHotword(AudioDataset):
+    # for finetuning contextual_paraformer and seaco_paraformer
+    def __init__(
+        self,
+        *args,
+        seaco_id: bool = 0,
+        **kwargs,
+    ):
+        """Initialize AudioDatasetHotword.
+        
+            Args:
+                *args: Variable positional arguments.
+                **kwargs: Additional keyword arguments.
+            """
+        super().__init__(*args, **kwargs)
+        self.seaco_id = seaco_id
+
+    def __getitem__(self, index):
+        """Internal: getitem  .
+        
+            Args:
+                index: TODO.
+            """
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+        if self.tokenizer:
+            ids = self.tokenizer.encode(target)
+            text = torch.tensor(ids, dtype=torch.int64)
+        else:
+            ids = target
+            text = ids
+        ids_lengths = len(ids)
+        text_lengths = torch.tensor([ids_lengths], dtype=torch.int32)
+
+        def generate_index(
+            length,
+            hotword_min_length=2,
+            hotword_max_length=8,
+            sample_rate=0.75,
+            double_rate=0.1,
+            pre_prob=0.0,
+            pre_index=None,
+            pre_hwlist=None,
+        ):
+            """Generate index.
+            
+                Args:
+                    length: TODO.
+                    hotword_min_length: TODO.
+                    hotword_max_length: TODO.
+                    sample_rate: TODO.
+                    double_rate: TODO.
+                    pre_prob: TODO.
+                    pre_index: TODO.
+                    pre_hwlist: TODO.
+                """
+            if length < hotword_min_length:
+                return [-1]
+            if random.random() < sample_rate:
+                if pre_prob > 0 and random.random() < pre_prob and pre_index is not None:
+                    return pre_index
+                if length == hotword_min_length:
+                    return [0, length - 1]
+                elif (
+                    random.random() < double_rate
+                    and length > hotword_max_length + hotword_min_length + 2
+                ):
+                    # sample two hotwords in a sentence
+                    _max_hw_length = min(hotword_max_length, length // 2)
+                    # first hotword
+                    start1 = random.randint(0, length // 3)
+                    end1 = random.randint(
+                        start1 + hotword_min_length - 1, start1 + _max_hw_length - 1
+                    )
+                    # second hotword
+                    start2 = random.randint(end1 + 1, length - hotword_min_length)
+                    end2 = random.randint(
+                        min(length - 1, start2 + hotword_min_length - 1),
+                        min(length - 1, start2 + hotword_max_length - 1),
+                    )
+                    return [start1, end1, start2, end2]
+                else:  # single hotword
+                    start = random.randint(0, length - hotword_min_length)
+                    end = random.randint(
+                        min(length - 1, start + hotword_min_length - 1),
+                        min(length - 1, start + hotword_max_length - 1),
+                    )
+                    return [start, end]
+            else:
+                return [-1]
+
+        hotword_indx = generate_index(text_lengths[0])
+        return {
+            "speech": speech[0, :, :],
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "hotword_indx": hotword_indx,
+            "seaco_id": self.seaco_id,
+        }
+
+    def collator(self, samples: list = None):
+        """Collator.
+        
+            Args:
+                samples: TODO.
+            """
+        outputs = {}
+        hotword_indxs = []
+        seaco_id = samples[0]["seaco_id"]
+        for sample in samples:
+            for key in sample.keys():
+                if key == "seaco_id":
+                    continue
+                elif key == "hotword_indx":
+                    hotword_indxs.append(sample[key])
+                else:
+                    if key not in outputs:
+                        outputs[key] = []
+                    outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+
+        hotword_list, hotword_lengths = [], []
+        text = outputs["text"]
+        seaco_label_pad = torch.ones_like(text) * -1 if seaco_id else None
+        for b, (hotword_indx, one_text, length) in enumerate(
+            zip(hotword_indxs, text, outputs["text_lengths"])
+        ):
+            length = length[0]
+            if seaco_label_pad is not None:
+                seaco_label_pad[b][:length] = seaco_id
+            if hotword_indx[0] != -1:
+                start, end = int(hotword_indx[0]), int(hotword_indx[1])
+                hotword = one_text[start : end + 1]
+                hotword_list.append(hotword)
+                hotword_lengths.append(end - start + 1)
+                if seaco_label_pad is not None:
+                    seaco_label_pad[b][start : end + 1] = one_text[start : end + 1]
+                if len(hotword_indx) == 4 and hotword_indx[2] != -1:
+                    # the second hotword if exist
+                    start, end = int(hotword_indx[2]), int(hotword_indx[3])
+                    hotword_list.append(one_text[start : end + 1])
+                    hotword_lengths.append(end - start + 1)
+                    if seaco_label_pad is not None:
+                        seaco_label_pad[b][start : end + 1] = one_text[start : end + 1]
+        hotword_list.append(torch.tensor([1]))
+        hotword_lengths.append(1)
+        hotword_pad = torch.nn.utils.rnn.pad_sequence(
+            hotword_list, batch_first=True, padding_value=0
+        )
+        outputs["hotword_pad"] = hotword_pad
+        outputs["hotword_lengths"] = torch.tensor(hotword_lengths, dtype=torch.int32)
+        if seaco_label_pad is not None:
+            outputs["seaco_label_pad"] = seaco_label_pad
+        return outputs

funasr/datasets/audio_datasets/espnet_samplers.py

@@ -0,0 +1,198 @@
+import torch
+import numpy as np
+import logging
+import math
+import torch.distributed as dist
+from torch.utils.data import DistributedSampler
+from torch.utils.data import BatchSampler, Sampler
+import torch.distributed as dist
+import random
+from funasr.register import tables
+
+
+@tables.register("batch_sampler_classes", "EspnetStyleBatchSampler")
+def EspnetStyleBatchSampler_fn(dataset, **kwargs):
+    """Espnetstylebatchsampler fn.
+    
+        Args:
+            dataset: TODO.
+            **kwargs: Additional keyword arguments.
+        """
+    dataloader_args = {}
+
+    batch_sampler = EspnetStyleBatchSampler(dataset, **kwargs)
+    dataloader_args["batch_sampler"] = batch_sampler
+    dataloader_args["num_workers"] = kwargs.get("num_workers", 4)
+    dataloader_args["pin_memory"] = kwargs.get("pin_memory", True)
+    num_workers = dataloader_args.get("num_workers", 4)
+    if num_workers > 0:
+        dataloader_args["persistent_workers"] = kwargs.get("persistent_workers", True)
+        dataloader_args["prefetch_factor"] = kwargs.get("prefetch_factor", 2)
+
+    return dataloader_args
+
+
+import torch
+from torch.utils.data import Dataset, DistributedSampler
+import math
+import random
+
+
+class EspnetStyleBatchSampler(DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        batch_type="token",
+        rank=None,
+        num_replicas=None,
+        rank_split=False,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        sort_size: int = 1024,
+        start_step: int = 0,
+        **kwargs,
+    ):
+
+        """Initialize EspnetStyleBatchSampler.
+        
+            Args:
+                dataset: TODO.
+                batch_size: Number of samples per batch.
+                batch_type: TODO.
+                rank: TODO.
+                num_replicas: TODO.
+                rank_split: TODO.
+                shuffle: TODO.
+                drop_last: TODO.
+                is_training: Boolean flag for training.
+                sort_size: Size/dimension parameter.
+                start_step: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        # if rank_split:
+        #     logging.info(f"Warning, rank_split: {rank_split}, batch and shuffle data in local rank")
+        #     rank = 0
+        #     num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.batch_type = batch_type
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+
+        self.total_size = len(self.dataset)
+        self.num_samples = int(math.ceil(self.total_size / self.num_replicas))
+        self.epoch = 0
+        self.sort_size = sort_size * num_replicas
+        self.max_token_length = kwargs.get("max_token_length", 2048)
+        self.min_token_length = kwargs.get("min_token_length", 0)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+        self.start_step = start_step
+        self.batch_num = 1
+        if self.start_step > 0:
+            logging.info(f"Warning, start_step > 0, dataloader start from step: {self.start_step}")
+        # super().__init__(dataset, num_replicas=num_replicas, rank=rank,
+        #                  shuffle=shuffle, drop_last=drop_last)
+
+    def __iter__(self):
+        """Internal: iter  ."""
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            random.seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Sort indices by sample length
+        sorted_indices = sorted(indices, key=lambda idx: self.dataset.get_source_len(idx))
+
+        # Organize batches based on 'length' or 'example'
+        buffer_batches = []
+        batch = []
+        max_len_in_batch = 0  # Tracks the max sample length within the current batch
+
+        for idx in sorted_indices:
+
+            # original_sample_length = self.dataset.get_source_len(idx)
+            # if (
+            #     original_sample_length < self.min_token_length
+            #     or original_sample_length > self.max_token_length
+            # ):  # Skip samples that exceed the max length
+            #     continue
+
+            # sample_length = 1 if self.batch_type == "example" else original_sample_length
+
+            # Set sample_length based on the batch type
+            if self.batch_type == "example":
+                sample_length = 1
+            elif self.batch_type == "token":
+                sample_length = self.dataset.get_source_len(idx) + int(
+                    self.dataset.get_target_len(idx) * 1.2
+                )
+            else:
+                sample_length = self.dataset.get_source_len(idx)
+            # Calculate potential batch size with the new sample
+            potential_batch_length = max(max_len_in_batch, sample_length) * (len(batch) + 1)
+            # Add index to batch if it doesn't exceed batch size limit
+            if potential_batch_length <= self.batch_size:
+                batch.append(idx)
+                max_len_in_batch = max(max_len_in_batch, sample_length)
+            else:
+                # Save the current batch and start a new one
+                buffer_batches.append(batch)
+                batch = [idx]
+                max_len_in_batch = sample_length
+
+        # Add the last batch if it shouldn't be dropped
+        if batch and (not self.drop_last or len(batch) * max_len_in_batch == self.batch_size):
+            buffer_batches.append(batch)
+
+        # Shuffle the list of batches
+        if self.shuffle:
+            random.seed(self.epoch)
+            random.shuffle(buffer_batches)
+
+        # Ensure each rank gets the same number of batches
+        batches_per_rank = int(math.ceil(len(buffer_batches) / self.num_replicas))
+        total_batches_needed = batches_per_rank * self.num_replicas
+        extra_batches = total_batches_needed - len(buffer_batches)
+        # Add extra batches by random selection, if needed
+        buffer_batches += random.choices(buffer_batches, k=extra_batches)
+
+        # Allocate the batches to the current rank
+        start_idx = self.rank * batches_per_rank
+        end_idx = start_idx + batches_per_rank
+        rank_batches = buffer_batches[start_idx + self.start_step : end_idx]
+
+        self.batch_num = len(rank_batches)
+
+        logging.info(
+            f"rank: {self.rank}, dataloader start from step: {self.start_step}, batch_num: {end_idx-start_idx}, batch_num_after_step: {len(rank_batches)}"
+        )
+        # Return an iterator over the batches for the current rank
+        return iter(rank_batches)
+
+    def __len__(self):
+        # Calculate the number of batches per epoch for the current rank
+        """Internal: len  ."""
+        return self.batch_num
+
+    def set_epoch(self, epoch):
+        # Set the epoch for shuffling
+        """Set epoch.
+        
+            Args:
+                epoch: TODO.
+            """
+        self.epoch = epoch

funasr/datasets/audio_datasets/index_ds.py

@@ -0,0 +1,173 @@
+import os
+import json
+import torch
+import logging
+
+import librosa
+import random
+import torch.distributed as dist
+
+from funasr.register import tables
+
+
+@tables.register("index_ds_classes", "IndexDSJsonl")
+@tables.register("index_ds_classes", "IndexDSJsonlRankFull")
+@tables.register("index_ds_classes", "IndexDSJsonlRankSplit")
+class IndexDSJsonlRankFull(torch.utils.data.Dataset):
+
+    def __init__(self, path: str, **kwargs):
+        """Initialize IndexDSJsonlRankFull.
+        
+            Args:
+                path: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        super().__init__()
+        self.max_source_length = kwargs.get("max_source_length", 2048)
+        self.min_source_length = kwargs.get("min_source_length", 0)
+        self.max_target_length = kwargs.get("max_target_length", 2048)
+        self.min_target_length = kwargs.get("min_target_length", 0)
+        self.max_token_length = kwargs.get("max_token_length", 2200)
+
+        is_training = kwargs.get("is_training", True)
+        if not (path.endswith(".jsonl") or path.endswith(".json")):
+            # jsonl list file
+            data_split_num = kwargs.get("data_split_num", 1)
+            data_split_i = kwargs.get("data_split_i", 0)
+
+            if not is_training:
+                data_split_num = 1
+                data_split_i = 0
+            with open(path, encoding="utf-8") as fin:
+                file_list_all = fin.readlines()
+
+                num_per_slice = (len(file_list_all) - 1) // data_split_num + 1  # 16
+                file_list = file_list_all[
+                    data_split_i * num_per_slice : (data_split_i + 1) * num_per_slice
+                ]
+                logging.info(
+                    f"is_training: {is_training}, data_split_num: {data_split_num}, data_split_i: {data_split_i}, \nfile_list: {file_list}, \nfile_list_all: {file_list_all}"
+                )
+
+        else:
+            file_list = [path]
+
+        # total_num = len(file_list)
+        # try:
+        #     rank = dist.get_rank()
+        #     world_size = dist.get_world_size()
+        # except:
+        #     rank = 0
+        #     world_size = 1
+        #     logging.info("distributed is not initialized, only single shard")
+        #
+        # if not kwargs.get("rank_split", False):
+        #     logging.info(f"Warning, rank_split disenabled, batch and shuffle data in global")
+        #     rank = 0
+        #     world_size = 1
+        #
+        # num_per_rank = total_num // world_size
+        # if num_per_rank * world_size < total_num:
+        #     logging.info(f"Warning, jsonl file:{total_num} could not be divided by world_size: {world_size}, {path}")
+        #     total_num_needed = num_per_rank * world_size
+        #
+        #     extra_num = total_num_needed - total_num
+        #     file_list_tmp = random.choices(file_list, k=extra_num)
+        #     file_list += file_list_tmp
+        #     logging.info(f"Warning, after random choices: {file_list}")
+        #
+        # file_list_rank = file_list[rank * num_per_rank:(rank + 1) * num_per_rank]
+        #
+        # logging.info(
+        #     f"is_training: {is_training}, file_list_rank: {file_list_rank}")
+
+        # contents = []
+        # for file_json in file_list_rank:
+        contents = []
+        for file_json in file_list:
+            with open(file_json.strip(), encoding="utf-8") as fin:
+                for line in fin:
+                    data = json.loads(line.strip())
+                    if "text" in data:  # for sft
+                        contents.append(data["text"])
+                    if "source" in data:  # for speech lab pretrain
+                        prompt = data.get("prompt", "<ASR>")
+                        source = data["source"].replace(
+                            "/cpfs01", "/cpfs_speech/data"
+                        )  # only use in alibaba gpu group: .replace("/cpfs01", "/cpfs_speech/data")
+                        target = data["target"]
+                        source_len = data.get("source_len", 1)
+                        target_len = data.get("target_len", 0)
+                        text_language = data.get("text_language", "")
+                        if "aishell" in source and text_language != "en":
+                            target = target.replace(" ", "")
+                        if (
+                            source_len < self.min_source_length
+                            or source_len > self.max_source_length
+                        ):
+                            continue
+                        if (
+                            target_len < self.min_target_length
+                            or target_len > self.max_target_length
+                        ):
+                            continue
+
+                        if (source_len + target_len) > self.max_token_length:
+                            continue
+
+                        contents_i = {
+                            "source": source,
+                            "prompt": prompt,
+                            "target": target,
+                            "source_len": source_len,
+                            "target_len": target_len,
+                        }
+                        text_language = data.get("text_language", None)
+                        if text_language is not None:
+                            contents_i["text_language"] = text_language
+                        if "emo_target" in data:
+                            contents_i["emo_target"] = data["emo_target"]
+                        if "event_target" in data:
+                            contents_i["event_target"] = data["event_target"]
+                        if "with_or_wo_itn" in data:
+                            contents_i["with_or_wo_itn"] = data["with_or_wo_itn"]
+                        # audio_language = data.get("audio_language", None)
+                        # if audio_language is not None:
+                        #     contents_i["audio_language"] = audio_language
+                        contents.append(contents_i)
+
+        self.contents = contents
+
+        logging.info("total_num of samplers: {}, {}".format(len(self.contents), path))
+
+    def __len__(self):
+        """Internal: len  ."""
+        return len(self.contents)
+
+    def __getitem__(self, index):
+
+        """Internal: getitem  .
+        
+            Args:
+                index: TODO.
+            """
+        data = self.contents[index]
+
+        return data
+
+    def get_source_len(self, data_dict):
+        """Get source len.
+        
+            Args:
+                data_dict: TODO.
+            """
+        return data_dict.get("source_len", 1)
+
+    def get_target_len(self, data_dict):
+
+        """Get target len.
+        
+            Args:
+                data_dict: TODO.
+            """
+        return data_dict.get("target_len", 0)

funasr/datasets/audio_datasets/jsonl2scp.py

@@ -0,0 +1,77 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+
+
+def gen_scp_from_jsonl(jsonl_file, data_type_list, wav_scp_file, text_file):
+
+    """Gen scp from jsonl.
+    
+        Args:
+            jsonl_file: TODO.
+            data_type_list: TODO.
+            wav_scp_file: TODO.
+            text_file: TODO.
+        """
+    wav_f = open(wav_scp_file, "w")
+    text_f = open(text_file, "w")
+    with open(jsonl_file, encoding="utf-8") as fin:
+        for line in fin:
+            data = json.loads(line.strip())
+
+            prompt = data.get("prompt", "<ASR>")
+            source = data[data_type_list[0]]
+            target = data[data_type_list[1]]
+            source_len = data.get("source_len", 1)
+            target_len = data.get("target_len", 0)
+            if "aishell" in source:
+                target = target.replace(" ", "")
+            key = data["key"]
+            wav_f.write(f"{key}\t{source}\n")
+            wav_f.flush()
+            text_f.write(f"{key}\t{target}\n")
+            text_f.flush()
+
+    wav_f.close()
+    text_f.close()
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    """Main hydra.
+    
+        Args:
+            cfg: Configuration overrides.
+        """
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get(
+        "scp_file_list",
+        ("/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"),
+    )
+    if isinstance(scp_file_list, str):
+        scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source", "target"))
+    jsonl_file = kwargs.get(
+        "jsonl_file_in", "/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl"
+    )
+    gen_scp_from_jsonl(jsonl_file, data_type_list, *scp_file_list)
+
+
+"""
+python -m funasr.datasets.audio_datasets.json2scp \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_in=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/preprocessor.py

@@ -0,0 +1,82 @@
+import os
+import json
+import torch
+import logging
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from typing import Collection
+import torch
+import torchaudio
+from torch import nn
+import random
+import re
+from funasr.tokenizer.cleaner import TextCleaner
+from funasr.register import tables
+
+
+@tables.register("preprocessor_classes", "SpeechPreprocessSpeedPerturb")
+class SpeechPreprocessSpeedPerturb(nn.Module):
+    def __init__(self, speed_perturb: list = None, **kwargs):
+        """Initialize SpeechPreprocessSpeedPerturb.
+        
+            Args:
+                speed_perturb: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        super().__init__()
+        self.speed_perturb = speed_perturb
+
+    def forward(self, waveform, fs, **kwargs):
+        """Forward pass for training.
+        
+            Args:
+                waveform: TODO.
+                fs: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        if self.speed_perturb is None:
+            return waveform
+        speed = random.choice(self.speed_perturb)
+        if speed != 1.0:
+            if not isinstance(waveform, torch.Tensor):
+                waveform = torch.tensor(waveform)
+            waveform, _ = torchaudio.sox_effects.apply_effects_tensor(
+                waveform.view(1, -1), fs, [["speed", str(speed)], ["rate", str(fs)]]
+            )
+            waveform = waveform.view(-1)
+
+        return waveform
+
+
+@tables.register("preprocessor_classes", "TextPreprocessSegDict")
+class TextPreprocessSegDict(nn.Module):
+    def __init__(
+        self,
+        seg_dict: str = None,
+        text_cleaner: Collection[str] = None,
+        split_with_space: bool = False,
+        **kwargs
+    ):
+        """Initialize TextPreprocessSegDict.
+        
+            Args:
+                seg_dict: TODO.
+                text_cleaner: TODO.
+                split_with_space: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        super().__init__()
+
+        self.text_cleaner = TextCleaner(text_cleaner)
+
+    def forward(self, text, **kwargs):
+        """Forward pass for training.
+        
+            Args:
+                text: Text tensor or string input.
+                **kwargs: Additional keyword arguments.
+            """
+        text = self.text_cleaner(text)
+
+        return text

funasr/datasets/audio_datasets/samplers.py

@@ -0,0 +1,592 @@
+import torch
+import numpy as np
+import logging
+import math
+import random
+import torch.distributed as dist
+from torch.utils.data import DistributedSampler
+from torch.utils.data import BatchSampler, Sampler
+import torch.distributed as dist
+
+from funasr.register import tables
+
+
+@tables.register("batch_sampler_classes", "BatchSampler")
+@tables.register("batch_sampler_classes", "CustomDistributedBatchSampler")
+@tables.register("batch_sampler_classes", "CustomDistributedDynamicBatchSampler")
+@tables.register("batch_sampler_classes", "DynamicBatchLocalShuffleSampler")
+@tables.register("batch_sampler_classes", "RankFullLocalShuffleBatchSampler")
+@tables.register("batch_sampler_classes", "RankFullLocalShuffleDynamicBatchSampler")
+def CustomDistributedBatchSampler_fn(dataset, **kwargs):
+    """Customdistributedbatchsampler fn.
+    
+        Args:
+            dataset: TODO.
+            **kwargs: Additional keyword arguments.
+        """
+    dataloader_args = {}
+    batch_type = kwargs.get("batch_type", "example")
+    if batch_type == "example":
+        batch_sampler = CustomDistributedBatchSampler(dataset, **kwargs)
+
+    else:
+        if kwargs.get("sort_size", -1) > 0:
+            batch_sampler = CustomDistributedBufferDynamicBatchSampler(dataset, **kwargs)
+        else:
+            batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)
+        # batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)
+
+    dataloader_args["batch_sampler"] = batch_sampler
+    dataloader_args["num_workers"] = kwargs.get("num_workers", 4)
+    dataloader_args["pin_memory"] = kwargs.get("pin_memory", True)
+    num_workers = dataloader_args.get("num_workers", 4)
+    if num_workers > 0:
+        dataloader_args["persistent_workers"] = kwargs.get("persistent_workers", True)
+        dataloader_args["prefetch_factor"] = kwargs.get("prefetch_factor", 2)
+
+    return dataloader_args
+
+
+class CustomDistributedBatchSampler(Sampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        **kwargs,
+    ):
+
+        """Initialize CustomDistributedBatchSampler.
+        
+            Args:
+                dataset: TODO.
+                batch_size: Number of samples per batch.
+                num_replicas: TODO.
+                rank: TODO.
+                shuffle: TODO.
+                drop_last: TODO.
+                is_training: Boolean flag for training.
+                **kwargs: Additional keyword arguments.
+            """
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+        # self.total_size = len(dataset)
+        if self.drop_last:
+            self.total_size = (len(self.dataset) // (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        else:
+            self.total_size = math.ceil(len(self.dataset) / (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        self.num_samples = int(self.total_size // self.num_replicas)
+        self.epoch = 0
+        self.max_token_length = kwargs.get("max_token_length", None)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+
+    def __iter__(self):
+        # Generate a list of indices
+        """Internal: iter  ."""
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Add extra samples to make it evenly divisible
+        padding_size = self.total_size - len(indices)
+        if padding_size <= len(indices):
+            indices += indices[:padding_size]
+        else:
+            indices += (
+                indices * (padding_size // len(indices)) + indices[: padding_size % len(indices)]
+            )
+
+        assert len(indices) == self.total_size
+
+        # Subsample
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        # Filter out indices with length greater than the max length, if provided
+        if self.max_token_length is not None:
+            filtered_indices = []
+            for idx in indices:
+                source_len = self.dataset.get_source_len(idx) / self.length_scale_source
+                if source_len <= self.max_token_length:
+                    filtered_indices.append(idx)
+            indices = filtered_indices
+
+        # Now that we have only the indices for this replica, chunk them into batches
+        batches = [
+            indices[i : i + self.batch_size] for i in range(0, len(indices), self.batch_size)
+        ]
+
+        # Drop the last batch if it's not full and drop_last is True
+        if self.drop_last and len(batches[-1]) != self.batch_size:
+            batches = batches[:-1]
+
+        return iter(batches)
+
+    def __len__(self):
+
+        """Internal: len  ."""
+        return self.num_samples // self.batch_size
+
+    def set_epoch(self, epoch):
+        """Set epoch.
+        
+            Args:
+                epoch: TODO.
+            """
+        self.epoch = epoch
+
+
+class CustomDistributedBufferBatchSampler(Sampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        sort_size: int = 1024,
+        **kwargs,
+    ):
+
+        """Initialize CustomDistributedBufferBatchSampler.
+        
+            Args:
+                dataset: TODO.
+                batch_size: Number of samples per batch.
+                num_replicas: TODO.
+                rank: TODO.
+                shuffle: TODO.
+                drop_last: TODO.
+                is_training: Boolean flag for training.
+                sort_size: Size/dimension parameter.
+                **kwargs: Additional keyword arguments.
+            """
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+        # self.total_size = len(dataset)
+        if self.drop_last:
+            self.total_size = (len(self.dataset) // (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        else:
+            self.total_size = math.ceil(len(self.dataset) / (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        self.num_samples = int(self.total_size // self.num_replicas)
+        self.epoch = 0
+        self.max_token_length = kwargs.get("max_token_length", None)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+        self.sort_size = sort_size
+
+    def __iter__(self):
+        # Generate a list of indices
+        """Internal: iter  ."""
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Add extra samples to make it evenly divisible
+        padding_size = self.total_size - len(indices)
+        if padding_size <= len(indices):
+            indices += indices[:padding_size]
+        else:
+            indices += (
+                indices * (padding_size // len(indices)) + indices[: padding_size % len(indices)]
+            )
+
+        assert len(indices) == self.total_size
+
+        # Subsample
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        # Filter out indices with length greater than the max length, if provided
+        if self.max_token_length is not None:
+            filtered_indices = []
+            for idx in indices:
+                source_len = self.dataset.get_source_len(idx) / self.length_scale_source
+                if source_len <= self.max_token_length:
+                    filtered_indices.append(idx)
+            indices = filtered_indices
+
+        # Buffer sorting logic
+        sorted_batches = []
+        buffer = []
+
+        for idx in indices:
+            buffer.append(idx)
+            if len(buffer) >= self.sort_size:
+                # Sort the buffer based on some criteria, e.g., dataset sample length
+                buffer.sort(key=lambda x: self.dataset.get_source_len(x))
+                sorted_batches.extend(self._create_batches_from_buffer(buffer))
+                buffer = []
+
+        # Handle the remaining items in the buffer
+        if buffer:
+            buffer.sort(key=lambda x: self.dataset.get_source_len(x))
+            sorted_batches.extend(self._create_batches_from_buffer(buffer))
+
+        return iter(sorted_batches)
+
+    def _create_batches_from_buffer(self, buffer):
+        # Function to convert the sorted buffer into batches
+        """Internal: create batches from buffer.
+        
+            Args:
+                buffer: TODO.
+            """
+        batched_buffer = [
+            buffer[i : i + self.batch_size] for i in range(0, len(buffer), self.batch_size)
+        ]
+        if self.drop_last and len(batched_buffer[-1]) != self.batch_size:
+            batched_buffer = batched_buffer[:-1]
+        return batched_buffer
+
+    def __len__(self):
+
+        """Internal: len  ."""
+        return self.num_samples // self.batch_size
+
+    def set_epoch(self, epoch):
+        """Set epoch.
+        
+            Args:
+                epoch: TODO.
+            """
+        self.epoch = epoch
+
+
+class CustomDistributedDynamicBatchSampler(DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        **kwargs,
+    ):
+
+        """Initialize CustomDistributedDynamicBatchSampler.
+        
+            Args:
+                dataset: TODO.
+                batch_size: Number of samples per batch.
+                num_replicas: TODO.
+                rank: TODO.
+                shuffle: TODO.
+                drop_last: TODO.
+                is_training: Boolean flag for training.
+                **kwargs: Additional keyword arguments.
+            """
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+
+        self.total_size = len(self.dataset)
+        # self.num_samples = int(math.ceil(self.total_size / self.num_replicas))
+        self.epoch = 0
+        self.max_token_length = kwargs.get("max_token_length", 2048)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+
+    def __iter__(self):
+        """Internal: iter  ."""
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+
+        batches = []
+        batch = []
+        max_len_in_batch = 0
+        current_batch_length = 0
+
+        for idx in indices:
+            sample_length = self.dataset.get_source_len(idx)
+            if sample_length > self.max_token_length:
+                continue
+            potential_batch_length = (
+                max_len_in_batch if sample_length < max_len_in_batch else sample_length
+            ) * (len(batch) + 1)
+
+            if potential_batch_length <= self.batch_size:
+                batch.append(idx)
+                if sample_length > max_len_in_batch:
+                    max_len_in_batch = sample_length
+                    # current_batch_length = max_len_in_batch * len(batch)
+            else:
+                batches.append(batch)
+                batch = [idx]
+                max_len_in_batch = sample_length
+                # current_batch_length = max_len_in_batch
+
+        # Add the last batch if it's not empty and we're not dropping it
+        if batch and (not self.drop_last or len(batch) * max_len_in_batch == self.batch_size):
+            batches.append(batch)
+
+        return iter(batches)
+
+    def __len__(self):
+
+        """Internal: len  ."""
+        return 1
+
+    def set_epoch(self, epoch):
+        """Set epoch.
+        
+            Args:
+                epoch: TODO.
+            """
+        self.epoch = epoch
+
+
+class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        batch_type="token",
+        num_replicas=None,
+        rank=None,
+        rank_split=False,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        sort_size: int = 1024,
+        start_step: int = 0,
+        **kwargs,
+    ):
+
+        """Initialize CustomDistributedBufferDynamicBatchSampler.
+        
+            Args:
+                dataset: TODO.
+                batch_size: Number of samples per batch.
+                batch_type: TODO.
+                num_replicas: TODO.
+                rank: TODO.
+                rank_split: TODO.
+                shuffle: TODO.
+                drop_last: TODO.
+                is_training: Boolean flag for training.
+                sort_size: Size/dimension parameter.
+                start_step: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+
+        # if rank_split:
+        #     logging.info(f"Warning, rank_split: {rank_split}, batch and shuffle data in local rank")
+        #     rank = 0
+        #     num_replicas = 1
+
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.batch_type = batch_type
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+
+        self.total_size = len(self.dataset)
+        self.num_samples = int(math.ceil(self.total_size / self.num_replicas))
+        self.epoch = 0
+        self.sort_size = sort_size * num_replicas
+        self.max_token_length = kwargs.get("max_token_length", 2048)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+        self.batch_size_sample_max = kwargs.get("batch_size_sample_max", 200)
+        self.start_step = start_step
+        self.batch_num = 1
+        if self.start_step > 0:
+            logging.info(f"Warning, start_step > 0, dataloader start from step: {self.start_step}")
+        # super().__init__(
+        #     dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle, drop_last=drop_last
+        # )
+
+    def __iter__(self):
+        """Internal: iter  ."""
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            random.seed(self.epoch)
+
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Create sorted buffers and form batches
+        buffer_batches = []
+        for i in range(0, len(indices), self.sort_size):
+            buffer = sorted(
+                indices[i : i + self.sort_size], key=lambda idx: self.dataset.get_source_len(idx)
+            )
+            batch = []
+            max_len_in_batch = 0
+            count = 1
+            for idx in buffer:
+                original_sample_length = self.dataset.get_source_len(idx)
+                if original_sample_length > self.max_token_length:
+                    continue
+                sample_length = 1 if self.batch_type == "example" else original_sample_length
+                potential_batch_length = max(max_len_in_batch, sample_length) * (len(batch) + 1)
+                if potential_batch_length <= self.batch_size and count < self.batch_size_sample_max:
+                    batch.append(idx)
+                    max_len_in_batch = max(max_len_in_batch, sample_length)
+                    count += 1
+                else:
+                    buffer_batches.append(batch)
+                    batch = [idx]
+                    max_len_in_batch = sample_length
+                    count = 1
+            if batch:
+                buffer_batches.append(batch)
+
+        # Ensure each rank gets the same number of batches, duplicate data if needed
+        batches_per_rank = math.ceil(len(buffer_batches) / self.num_replicas)
+        total_batches_needed = batches_per_rank * self.num_replicas
+
+        extra_batches = total_batches_needed - len(buffer_batches)
+        buffer_batches += random.choices(buffer_batches, k=extra_batches)
+
+        # Evenly distribute batches from buffer_batches to each rank
+        rank_batches = [[] for _ in range(self.num_replicas)]
+        for i, batch in enumerate(buffer_batches):
+            rank_batches[i % self.num_replicas].append(batch)
+
+        # Assign all batches for the current rank directly
+        final_batches = rank_batches[self.rank][self.start_step :]
+        self.batch_num = len(final_batches)
+
+        logging.info(
+            f"rank: {self.rank}, dataloader start from step: {self.start_step}, batch_num: {len(rank_batches[self.rank])}, after: {self.batch_num}"
+        )
+        return iter(final_batches)
+
+    def __len__(self):
+        # Calculate the number of batches per epoch for the current rank
+        """Internal: len  ."""
+        return self.batch_num
+
+    def set_epoch(self, epoch):
+        """Set epoch.
+        
+            Args:
+                epoch: TODO.
+            """
+        self.epoch = epoch
+
+
+class DistributedSamplerWarp(BatchSampler):
+    def __init__(
+        self, dataset, batch_size, num_replicas=None, rank=None, shuffle=True, drop_last=False
+    ):
+        """Initialize DistributedSamplerWarp.
+        
+            Args:
+                dataset: TODO.
+                batch_size: Number of samples per batch.
+                num_replicas: TODO.
+                rank: TODO.
+                shuffle: TODO.
+                drop_last: TODO.
+            """
+        if num_replicas is None:
+            if not torch.distributed.is_available():
+                raise RuntimeError("Requires distributed package to be available")
+            num_replicas = torch.distributed.get_world_size()
+        if rank is None:
+            if not torch.distributed.is_available():
+                raise RuntimeError("Requires distributed package to be available")
+            rank = torch.distributed.get_rank()
+
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.shuffle = shuffle
+        self.drop_last = drop_last
+
+        # Create an instance of the DistributedSampler
+        self.sampler = DistributedSampler(
+            self.dataset, num_replicas=self.num_replicas, rank=self.rank, shuffle=self.shuffle
+        )
+
+        # Call BatchSampler's constructor
+        super().__init__(self.sampler, batch_size, drop_last)
+
+    def __iter__(self):
+        # If we shuffle, we need to call the set_epoch method
+        """Internal: iter  ."""
+        if self.shuffle:
+            self.sampler.set_epoch(self.epoch)
+
+        # Generate batch indices using the parent class
+        return super().__iter__()
+
+    def set_epoch(self, epoch):
+        """Set epoch.
+        
+            Args:
+                epoch: TODO.
+            """
+        self.epoch = epoch

funasr/datasets/audio_datasets/scp2jsonl.py

@@ -0,0 +1,148 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from tqdm import tqdm
+
+
+def gen_jsonl_from_wav_text_list(
+    path, data_type_list=("source", "target"), jsonl_file_out: str = None, **kwargs
+):
+    """Gen jsonl from wav text list.
+    
+        Args:
+            path: TODO.
+            data_type_list: TODO.
+            jsonl_file_out: TODO.
+            **kwargs: Additional keyword arguments.
+        """
+    try:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    except:
+        rank = 0
+        world_size = 1
+
+    cpu_cores = os.cpu_count() or 1
+    print(f"convert wav.scp text to jsonl, ncpu: {cpu_cores}")
+    if rank == 0:
+        json_dict = {}
+        for data_type, data_file in zip(data_type_list, path):
+            json_dict[data_type] = {}
+            with open(data_file, "r") as f:
+
+                data_file_lists = f.readlines()
+                lines_for_each_th = (len(data_file_lists) - 1) // cpu_cores + 1
+                task_num = cpu_cores if len(data_file_lists) > cpu_cores else 1
+                # import pdb;pdb.set_trace()
+                if task_num > 1:
+                    with concurrent.futures.ThreadPoolExecutor(max_workers=cpu_cores) as executor:
+
+                        futures = [
+                            executor.submit(
+                                parse_context_length,
+                                data_file_lists[
+                                    i * lines_for_each_th : (i + 1) * lines_for_each_th
+                                ],
+                                data_type,
+                                i,
+                            )
+                            for i in range(task_num)
+                        ]
+
+                        for future in concurrent.futures.as_completed(futures):
+
+                            json_dict[data_type].update(future.result())
+                else:
+                    res = parse_context_length(data_file_lists, data_type)
+                    json_dict[data_type].update(res)
+
+        with open(jsonl_file_out, "w") as f:
+            for key in json_dict[data_type_list[0]].keys():
+                jsonl_line = {"key": key}
+                for data_file in data_type_list:
+                    if key in json_dict[data_file]:
+                        jsonl_line.update(json_dict[data_file][key])
+                jsonl_line = json.dumps(jsonl_line, ensure_ascii=False)
+                f.write(jsonl_line + "\n")
+                f.flush()
+        print(f"processed {len(json_dict[data_type_list[0]])} samples")
+
+    else:
+        pass
+
+    if world_size > 1:
+        dist.barrier()
+
+
+def parse_context_length(data_list: list, data_type: str, id=0):
+    """Parse context length.
+    
+        Args:
+            data_list: TODO.
+            data_type: TODO.
+            id: TODO.
+        """
+    pbar = tqdm(total=len(data_list), dynamic_ncols=True)
+    res = {}
+    for i, line in enumerate(data_list):
+        pbar.update(1)
+        pbar.set_description(f"cpu: {id}")
+        lines = line.strip().split(maxsplit=1)
+        key = lines[0]
+        line = lines[1] if len(lines) > 1 else ""
+        line = line.strip()
+        if data_type == "source":
+            if os.path.exists(line):
+                waveform, _ = librosa.load(line, sr=16000)
+                sample_num = len(waveform)
+                context_len = int(sample_num * 1000 / 16000 / 10)
+            else:
+                print("source file not found: {}".format(line))
+                continue
+        else:
+            context_len = len(line.split()) if " " in line else len(line)
+        res[key] = {data_type: line, f"{data_type}_len": context_len}
+    return res
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    """Main hydra.
+    
+        Args:
+            cfg: Configuration overrides.
+        """
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get(
+        "scp_file_list",
+        ("/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"),
+    )
+    if isinstance(scp_file_list, str):
+        scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source", "target"))
+    jsonl_file_out = kwargs.get(
+        "jsonl_file_out", "/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl"
+    )
+    gen_jsonl_from_wav_text_list(
+        scp_file_list, data_type_list=data_type_list, jsonl_file_out=jsonl_file_out
+    )
+
+
+"""
+python -m funasr.datasets.audio_datasets.scp2jsonl \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_out=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/scp2len.py

@@ -0,0 +1,141 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from tqdm import tqdm
+
+
+def gen_jsonl_from_wav_text_list(
+    path, data_type_list=("source",), jsonl_file_out: str = None, **kwargs
+):
+    """Gen jsonl from wav text list.
+    
+        Args:
+            path: TODO.
+            data_type_list: TODO.
+            jsonl_file_out: TODO.
+            **kwargs: Additional keyword arguments.
+        """
+    try:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    except:
+        rank = 0
+        world_size = 1
+
+    cpu_cores = os.cpu_count() or 1
+    print(f"convert wav.scp text to jsonl, ncpu: {cpu_cores}")
+    if rank == 0:
+        json_dict = {}
+        # for data_type, data_file in zip(data_type_list, path):
+        data_type = data_type_list[0]
+        data_file = path
+        json_dict[data_type] = {}
+        with open(data_file, "r") as f:
+
+            data_file_lists = f.readlines()
+            print("")
+            lines_for_each_th = (len(data_file_lists) - 1) // cpu_cores + 1
+            task_num = cpu_cores if len(data_file_lists) > cpu_cores else 1
+            # import pdb;pdb.set_trace()
+            if task_num > 1:
+                with concurrent.futures.ThreadPoolExecutor(max_workers=cpu_cores) as executor:
+
+                    futures = [
+                        executor.submit(
+                            parse_context_length,
+                            data_file_lists[i * lines_for_each_th : (i + 1) * lines_for_each_th],
+                            data_type,
+                            i,
+                        )
+                        for i in range(task_num)
+                    ]
+
+                    for future in concurrent.futures.as_completed(futures):
+
+                        json_dict[data_type].update(future.result())
+            else:
+                res = parse_context_length(data_file_lists, data_type)
+                json_dict[data_type].update(res)
+
+        with open(jsonl_file_out, "w") as f:
+            for key in json_dict[data_type_list[0]].keys():
+                jsonl_line = {"key": key}
+                for data_file in data_type_list:
+                    jsonl_line.update(json_dict[data_file][key])
+                # jsonl_line = json.dumps(jsonl_line, ensure_ascii=False)
+                source_len = jsonl_line["source_len"]
+                jsonl_line = f"{key} {source_len}"
+                f.write(jsonl_line + "\n")
+                f.flush()
+        print(f"processed {len(json_dict[data_type_list[0]])} samples")
+
+    else:
+        pass
+
+    if world_size > 1:
+        dist.barrier()
+
+
+def parse_context_length(data_list: list, data_type: str, id=0):
+    """Parse context length.
+    
+        Args:
+            data_list: TODO.
+            data_type: TODO.
+            id: TODO.
+        """
+    pbar = tqdm(total=len(data_list), dynamic_ncols=True)
+    res = {}
+    for i, line in enumerate(data_list):
+        pbar.update(1)
+        pbar.set_description(f"cpu: {id}")
+        lines = line.strip().split(maxsplit=1)
+        key = lines[0]
+        line = lines[1] if len(lines) > 1 else ""
+        line = line.strip()
+        if os.path.exists(line):
+            waveform, _ = librosa.load(line, sr=16000)
+            sample_num = len(waveform)
+            context_len = int(sample_num / 16000 * 1000 / 10)
+        else:
+            context_len = len(line.split()) if " " in line else len(line)
+        res[key] = {data_type: line, f"{data_type}_len": context_len}
+    return res
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    """Main hydra.
+    
+        Args:
+            cfg: Configuration overrides.
+        """
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get("scp_file_list", "/Users/zhifu/funasr1.0/data/list/train_wav.scp")
+    # if isinstance(scp_file_list, str):
+    #     scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source",))
+    jsonl_file_out = kwargs.get("jsonl_file_out", "/Users/zhifu/funasr1.0/data/list/wav_len.txt")
+    gen_jsonl_from_wav_text_list(
+        scp_file_list, data_type_list=data_type_list, jsonl_file_out=jsonl_file_out
+    )
+
+
+"""
+python -m funasr.datasets.audio_datasets.scp2jsonl \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_out=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/sensevoice2jsonl.py

@@ -0,0 +1,216 @@
+import os
+import json
+import torch
+import logging
+import hydra
+import re
+import string
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from tqdm import tqdm
+
+
+def gen_jsonl_from_wav_text_list(
+    path, data_type_list=("source", "target"), jsonl_file_out: str = None, model_dir: str = "iic/SenseVoiceSmall", **kwargs
+):
+    """Gen jsonl from wav text list.
+    
+        Args:
+            path: TODO.
+            data_type_list: TODO.
+            jsonl_file_out: TODO.
+            model_dir: TODO.
+            **kwargs: Additional keyword arguments.
+        """
+    try:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    except:
+        rank = 0
+        world_size = 1
+
+    cpu_cores = os.cpu_count() or 1
+    print(f"convert wav.scp text to jsonl, ncpu: {cpu_cores}")
+    if rank == 0:
+        json_dict = {}
+        for data_type, data_file in zip(data_type_list, path):
+            json_dict[data_type] = {}
+            with open(data_file, "r") as f:
+
+                data_file_lists = f.readlines()
+                lines_for_each_th = (len(data_file_lists) - 1) // cpu_cores + 1
+                task_num = cpu_cores if len(data_file_lists) > cpu_cores else 1
+                # import pdb;pdb.set_trace()
+                if task_num > 1:
+                    with concurrent.futures.ThreadPoolExecutor(max_workers=cpu_cores) as executor:
+
+                        futures = [
+                            executor.submit(
+                                parse_context_length,
+                                data_file_lists[
+                                    i * lines_for_each_th : (i + 1) * lines_for_each_th
+                                ],
+                                data_type,
+                                i,
+                            )
+                            for i in range(task_num)
+                        ]
+
+                        for future in concurrent.futures.as_completed(futures):
+
+                            json_dict[data_type].update(future.result())
+                else:
+                    res = parse_context_length(data_file_lists, data_type)
+                    json_dict[data_type].update(res)
+
+        if "text_language" not in data_type_list or "emo_target" not in data_type_list or "event_target" not in data_type_list:
+            from funasr import AutoModel
+
+            model = AutoModel(
+                model=model_dir,
+            )
+
+            rich_dict = {}
+            for key in json_dict["source"].keys():
+                input_wav = json_dict["source"][key]["source"]
+                res = model.generate(
+                input=input_wav,
+                cache={},
+                language="auto",  # "zn", "en", "yue", "ja", "ko", "nospeech"
+                use_itn=True,
+                )
+                text = res[0]["text"]
+                pattern = r"<\|[^|]+\|>"
+                matches = re.findall(pattern, text)
+                text_language, emo_target, event_target = matches[:3]
+                rich_dict[key] = [text_language, emo_target, event_target]
+
+
+            if "text_language" not in data_type_list:
+                data_type_list.append("text_language")
+                if "text_language" not in json_dict:
+                    json_dict["text_language"] = {}
+                for key in json_dict["source"].keys():
+                    json_dict["text_language"][key] = {}
+                    json_dict["text_language"][key]["text_language"] = rich_dict[key][0]
+
+            if "emo_target" not in data_type_list:
+                data_type_list.append("emo_target")
+                if "emo_target" not in json_dict:
+                    json_dict["emo_target"] = {}
+                for key in json_dict["source"].keys():
+                    json_dict["emo_target"][key] = {}
+                    json_dict["emo_target"][key]["emo_target"] = rich_dict[key][1]
+
+            if "event_target" not in data_type_list:
+                data_type_list.append("event_target")
+                if "event_target" not in json_dict:
+                    json_dict["event_target"] = {}
+                for key in json_dict["source"].keys():
+                    json_dict["event_target"][key] = {}
+                    json_dict["event_target"][key]["event_target"] = rich_dict[key][2]
+
+        with open(jsonl_file_out, "w") as f:
+            for key in json_dict[data_type_list[0]].keys():
+                jsonl_line = {"key": key}
+                for data_file in data_type_list:
+                    jsonl_line.update(json_dict[data_file][key])
+                jsonl_line = json.dumps(jsonl_line, ensure_ascii=False)
+                f.write(jsonl_line + "\n")
+                f.flush()
+        print(f"processed {len(json_dict[data_type_list[0]])} samples")
+
+    else:
+        pass
+
+    if world_size > 1:
+        dist.barrier()
+
+def contains_punctuation(s):
+    """Contains punctuation.
+    
+        Args:
+            s: TODO.
+        """
+    punctuations = (
+        string.punctuation +  
+        '，。、；：？！""''（）【】《》〈〉「」『』〔〕［］｛｝～·…—–'  
+    )
+    return any(char in punctuations for char in s)
+
+def parse_context_length(data_list: list, data_type: str, id=0):
+    """Parse context length.
+    
+        Args:
+            data_list: TODO.
+            data_type: TODO.
+            id: TODO.
+        """
+    pbar = tqdm(total=len(data_list), dynamic_ncols=True)
+    res = {}
+    for i, line in enumerate(data_list):
+        pbar.update(1)
+        pbar.set_description(f"cpu: {id}")
+        lines = line.strip().split(maxsplit=1)
+        key = lines[0]
+        line = lines[1] if len(lines) > 1 else ""
+        line = line.strip()
+        if os.path.exists(line):
+            waveform, _ = librosa.load(line, sr=16000)
+            sample_num = len(waveform)
+            context_len = int(sample_num / 16000 * 1000 / 10)
+        else:
+            context_len = len(line.split()) if " " in line else len(line)
+        if data_type == "source":
+            res[key] = {data_type: line, f"{data_type}_len": context_len}
+        elif data_type == "target":
+            punc = contains_punctuation(line)
+            if punc:
+                with_or_wo_itn = "<|withitn|>"
+            else:
+                with_or_wo_itn = "<|woitn|>"
+            res[key] = {data_type: line, f"{data_type}_len": context_len, "with_or_wo_itn": with_or_wo_itn}
+        else:
+            res[key] = {data_type: line}
+    return res
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    """Main hydra.
+    
+        Args:
+            cfg: Configuration overrides.
+        """
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get(
+        "scp_file_list",
+        ("/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"),
+    )
+    if isinstance(scp_file_list, str):
+        scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source", "target"))
+    jsonl_file_out = kwargs.get(
+        "jsonl_file_out", "/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl"
+    )
+    model_dir = kwargs.get("model_dir", "iic/SenseVoiceSmall")
+    gen_jsonl_from_wav_text_list(
+        scp_file_list, data_type_list=data_type_list, jsonl_file_out=jsonl_file_out, model_dir=model_dir
+    )
+
+
+"""
+python -m funasr.datasets.audio_datasets.sensevoice2jsonl \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt", "/Users/zhifu/funasr1.0/test_local/text_language.txt", "/Users/zhifu/funasr1.0/test_local/emo_target.txt", "/Users/zhifu/funasr1.0/test_local/event_target.txt"]' \
+++data_type_list='["source", "target", "text_language", "emo_target", "event_target"]' \
+++jsonl_file_out='/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl' \
+++model_dir='iic/SenseVoiceSmall'
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/update_jsonl.py

@@ -0,0 +1,124 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+import threading
+from tqdm import tqdm
+from concurrent.futures import ThreadPoolExecutor
+
+
+def gen_scp_from_jsonl(jsonl_file, jsonl_file_out, ncpu):
+    """Gen scp from jsonl.
+    
+        Args:
+            jsonl_file: TODO.
+            jsonl_file_out: TODO.
+            ncpu: TODO.
+        """
+    jsonl_file_out_f = open(jsonl_file_out, "w")
+    with open(jsonl_file, encoding="utf-8") as fin:
+        lines = fin.readlines()
+
+        num_total = len(lines)
+        if ncpu > 1:
+            # 使用ThreadPoolExecutor限制并发线程数
+            with ThreadPoolExecutor(max_workers=ncpu) as executor:
+                # 提交任务到线程池
+                futures = {executor.submit(update_data, lines, i) for i in tqdm(range(num_total))}
+
+                # 等待所有任务完成，这会阻塞直到所有提交的任务完成
+                for future in concurrent.futures.as_completed(futures):
+                    # 这里可以添加额外的逻辑来处理完成的任务，但在这个例子中我们只是等待
+                    pass
+        else:
+            for i in range(num_total):
+                update_data(lines, i)
+        logging.info("All audio durations have been processed.")
+
+        for line in lines:
+
+            jsonl_file_out_f.write(line + "\n")
+            jsonl_file_out_f.flush()
+
+    jsonl_file_out_f.close()
+
+
+def update_data(lines, i):
+    """Update data.
+    
+        Args:
+            lines: TODO.
+            i: TODO.
+        """
+    line = lines[i]
+    data = json.loads(line.strip())
+
+    wav_path = data["source"].replace("/cpfs01", "/cpfs_speech/data")
+    if os.path.exists(wav_path):
+        waveform, _ = librosa.load(wav_path, sr=16000)
+        sample_num = len(waveform)
+        source_len = int(sample_num / 16000 * 1000 / 10)
+        source_len_old = data["source_len"]
+        # if (source_len_old - source_len) > 100 or (source_len - source_len_old) > 100:
+        #     logging.info(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
+        data["source_len"] = source_len
+        data["source"] = wav_path
+        jsonl_line = json.dumps(data, ensure_ascii=False)
+        lines[i] = jsonl_line
+
+
+def update_wav_len(jsonl_file_list_in, jsonl_file_out_dir, ncpu=1):
+
+    """Update wav len.
+    
+        Args:
+            jsonl_file_list_in: TODO.
+            jsonl_file_out_dir: TODO.
+            ncpu: TODO.
+        """
+    os.makedirs(jsonl_file_out_dir, exist_ok=True)
+    with open(jsonl_file_list_in, "r") as f:
+        data_file_lists = f.readlines()
+
+        for i, jsonl in enumerate(data_file_lists):
+            filename_with_extension = os.path.basename(jsonl.strip())
+            jsonl_file_out = os.path.join(jsonl_file_out_dir, filename_with_extension)
+            logging.info(f"{i}/{len(data_file_lists)}, jsonl: {jsonl}, {jsonl_file_out}")
+
+            gen_scp_from_jsonl(jsonl.strip(), jsonl_file_out, ncpu)
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    """Main hydra.
+    
+        Args:
+            cfg: Configuration overrides.
+        """
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    logging.info(kwargs)
+
+    jsonl_file_list_in = kwargs.get(
+        "jsonl_file_list_in", "/Users/zhifu/funasr1.0/data/list/data_jsonl.list"
+    )
+    jsonl_file_out_dir = kwargs.get("jsonl_file_out_dir", "/Users/zhifu/funasr1.0/data_tmp")
+    ncpu = kwargs.get("ncpu", 1)
+    update_wav_len(jsonl_file_list_in, jsonl_file_out_dir, ncpu)
+    # gen_scp_from_jsonl(jsonl_file_list_in, jsonl_file_out_dir)
+
+
+"""
+python -m funasr.datasets.audio_datasets.json2scp \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_in=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()