Initial commit: FunASR Speech Recognition Toolkit

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

funasr/models/lcbnet/attention.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# Copyright 2024 yufan
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Multi-Head Attention Return Weight layer definition."""
+
+import math
+
+import torch
+from torch import nn
+
+
+class MultiHeadedAttentionReturnWeight(nn.Module):
+    """Multi-Head Attention layer.
+
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, n_head, n_feat, dropout_rate):
+        """Construct an MultiHeadedAttentionReturnWeight object."""
+        super(MultiHeadedAttentionReturnWeight, self).__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        self.linear_q = nn.Linear(n_feat, n_feat)
+        self.linear_k = nn.Linear(n_feat, n_feat)
+        self.linear_v = nn.Linear(n_feat, n_feat)
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.attn = None
+        self.dropout = nn.Dropout(p=dropout_rate)
+
+    def forward_qkv(self, query, key, value):
+        """Transform query, key and value.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+
+        Returns:
+            torch.Tensor: Transformed query tensor (#batch, n_head, time1, d_k).
+            torch.Tensor: Transformed key tensor (#batch, n_head, time2, d_k).
+            torch.Tensor: Transformed value tensor (#batch, n_head, time2, d_k).
+
+        """
+        n_batch = query.size(0)
+        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
+        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
+        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
+        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
+        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
+        v = v.transpose(1, 2)  # (batch, head, time2, d_k)
+
+        return q, k, v
+
+    def forward_attention(self, value, scores, mask):
+        """Compute attention context vector.
+
+        Args:
+            value (torch.Tensor): Transformed value (#batch, n_head, time2, d_k).
+            scores (torch.Tensor): Attention score (#batch, n_head, time1, time2).
+            mask (torch.Tensor): Mask (#batch, 1, time2) or (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
+
+        """
+        n_batch = value.size(0)
+        if mask is not None:
+            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+            min_value = torch.finfo(scores.dtype).min
+            scores = scores.masked_fill(mask, min_value)
+            attn = torch.softmax(scores, dim=-1).masked_fill(
+                mask, 0.0
+            )  # (batch, head, time1, time2)
+        else:
+            attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        p_attn = self.dropout(attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (
+            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
+        )  # (batch, time1, d_model)
+
+        return self.linear_out(x), attn  # (batch, time1, d_model)
+
+    def forward(self, query, key, value, mask):
+        """Compute scaled dot product attention.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+        return self.forward_attention(v, scores, mask)

funasr/models/lcbnet/encoder.py

@@ -0,0 +1,430 @@
+# Copyright 2019 Shigeki Karita
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Transformer encoder definition."""
+
+from typing import List
+from typing import Optional
+from typing import Tuple
+
+import torch
+from torch import nn
+import logging
+
+from funasr.models.transformer.attention import MultiHeadedAttention
+from funasr.models.lcbnet.attention import MultiHeadedAttentionReturnWeight
+from funasr.models.transformer.embedding import PositionalEncoding
+from funasr.models.transformer.layer_norm import LayerNorm
+
+from funasr.models.transformer.utils.nets_utils import make_pad_mask
+from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward
+from funasr.models.transformer.utils.repeat import repeat
+from funasr.register import tables
+
+
+class EncoderLayer(nn.Module):
+    """Encoder layer module.
+
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention` instance
+            can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
+            can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool): Whether to use layer_norm before the first block.
+        concat_after (bool): Whether to concat attention layer's input and output.
+            if True, additional linear will be applied.
+            i.e. x -> x + linear(concat(x, att(x)))
+            if False, no additional linear will be applied. i.e. x -> x + att(x)
+        stochastic_depth_rate (float): Proability to skip this layer.
+            During training, the layer may skip residual computation and return input
+            as-is with given probability.
+    """
+
+    def __init__(
+        self,
+        size,
+        self_attn,
+        feed_forward,
+        dropout_rate,
+        normalize_before=True,
+        concat_after=False,
+        stochastic_depth_rate=0.0,
+    ):
+        """Construct an EncoderLayer object."""
+        super(EncoderLayer, self).__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.norm1 = LayerNorm(size)
+        self.norm2 = LayerNorm(size)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+        self.concat_after = concat_after
+        if self.concat_after:
+            self.concat_linear = nn.Linear(size + size, size)
+        self.stochastic_depth_rate = stochastic_depth_rate
+
+    def forward(self, x, mask, cache=None):
+        """Compute encoded features.
+
+        Args:
+            x_input (torch.Tensor): Input tensor (#batch, time, size).
+            mask (torch.Tensor): Mask tensor for the input (#batch, time).
+            cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time).
+
+        """
+        skip_layer = False
+        # with stochastic depth, residual connection `x + f(x)` becomes
+        # `x <- x + 1 / (1 - p) * f(x)` at training time.
+        stoch_layer_coeff = 1.0
+        if self.training and self.stochastic_depth_rate > 0:
+            skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
+            stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)
+
+        if skip_layer:
+            if cache is not None:
+                x = torch.cat([cache, x], dim=1)
+            return x, mask
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm1(x)
+
+        if cache is None:
+            x_q = x
+        else:
+            assert cache.shape == (x.shape[0], x.shape[1] - 1, self.size)
+            x_q = x[:, -1:, :]
+            residual = residual[:, -1:, :]
+            mask = None if mask is None else mask[:, -1:, :]
+
+        if self.concat_after:
+            x_concat = torch.cat((x, self.self_attn(x_q, x, x, mask)), dim=-1)
+            x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
+        else:
+            x = residual + stoch_layer_coeff * self.dropout(self.self_attn(x_q, x, x, mask))
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        if cache is not None:
+            x = torch.cat([cache, x], dim=1)
+
+        return x, mask
+
+
+@tables.register("encoder_classes", "TransformerTextEncoder")
+class TransformerTextEncoder(nn.Module):
+    """Transformer text encoder module.
+
+    Args:
+        input_size: input dim
+        output_size: dimension of attention
+        attention_heads: the number of heads of multi head attention
+        linear_units: the number of units of position-wise feed forward
+        num_blocks: the number of decoder blocks
+        dropout_rate: dropout rate
+        attention_dropout_rate: dropout rate in attention
+        positional_dropout_rate: dropout rate after adding positional encoding
+        input_layer: input layer type
+        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
+        normalize_before: whether to use layer_norm before the first block
+        concat_after: whether to concat attention layer's input and output
+            if True, additional linear will be applied.
+            i.e. x -> x + linear(concat(x, att(x)))
+            if False, no additional linear will be applied.
+            i.e. x -> x + att(x)
+        positionwise_layer_type: linear of conv1d
+        positionwise_conv_kernel_size: kernel size of positionwise conv1d layer
+        padding_idx: padding_idx for input_layer=embed
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        pos_enc_class=PositionalEncoding,
+        normalize_before: bool = True,
+        concat_after: bool = False,
+    ):
+        """Initialize TransformerTextEncoder.
+        
+            Args:
+                input_size: Size/dimension parameter.
+                output_size: Size/dimension parameter.
+                attention_heads: TODO.
+                linear_units: TODO.
+                num_blocks: TODO.
+                dropout_rate: TODO.
+                positional_dropout_rate: TODO.
+                attention_dropout_rate: TODO.
+                pos_enc_class: TODO.
+                normalize_before: TODO.
+                concat_after: TODO.
+            """
+        super().__init__()
+        self._output_size = output_size
+
+        self.embed = torch.nn.Sequential(
+            torch.nn.Embedding(input_size, output_size),
+            pos_enc_class(output_size, positional_dropout_rate),
+        )
+
+        self.normalize_before = normalize_before
+
+        positionwise_layer = PositionwiseFeedForward
+        positionwise_layer_args = (
+            output_size,
+            linear_units,
+            dropout_rate,
+        )
+        self.encoders = repeat(
+            num_blocks,
+            lambda lnum: EncoderLayer(
+                output_size,
+                MultiHeadedAttention(attention_heads, output_size, attention_dropout_rate),
+                positionwise_layer(*positionwise_layer_args),
+                dropout_rate,
+                normalize_before,
+                concat_after,
+            ),
+        )
+        if self.normalize_before:
+            self.after_norm = LayerNorm(output_size)
+
+    def output_size(self) -> int:
+        """Output size."""
+        return self._output_size
+
+    def forward(
+        self,
+        xs_pad: torch.Tensor,
+        ilens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """Embed positions in tensor.
+
+        Args:
+            xs_pad: input tensor (B, L, D)
+            ilens: input length (B)
+        Returns:
+            position embedded tensor and mask
+        """
+        masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
+        xs_pad = self.embed(xs_pad)
+
+        xs_pad, masks = self.encoders(xs_pad, masks)
+
+        if self.normalize_before:
+            xs_pad = self.after_norm(xs_pad)
+
+        olens = masks.squeeze(1).sum(1)
+        return xs_pad, olens, None
+
+
+@tables.register("encoder_classes", "FusionSANEncoder")
+class SelfSrcAttention(nn.Module):
+    """Single decoder layer module.
+
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` instance can be used as the argument.
+        src_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` instance can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
+            can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool): Whether to use layer_norm before the first block.
+        concat_after (bool): Whether to concat attention layer's input and output.
+            if True, additional linear will be applied.
+            i.e. x -> x + linear(concat(x, att(x)))
+            if False, no additional linear will be applied. i.e. x -> x + att(x)
+
+
+    """
+
+    def __init__(
+        self,
+        size,
+        attention_heads,
+        attention_dim,
+        linear_units,
+        self_attention_dropout_rate,
+        src_attention_dropout_rate,
+        positional_dropout_rate,
+        dropout_rate,
+        normalize_before=True,
+        concat_after=False,
+    ):
+        """Construct an SelfSrcAttention object."""
+        super(SelfSrcAttention, self).__init__()
+        self.size = size
+        self.self_attn = MultiHeadedAttention(
+            attention_heads, attention_dim, self_attention_dropout_rate
+        )
+        self.src_attn = MultiHeadedAttentionReturnWeight(
+            attention_heads, attention_dim, src_attention_dropout_rate
+        )
+        self.feed_forward = PositionwiseFeedForward(
+            attention_dim, linear_units, positional_dropout_rate
+        )
+        self.norm1 = LayerNorm(size)
+        self.norm2 = LayerNorm(size)
+        self.norm3 = LayerNorm(size)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.normalize_before = normalize_before
+        self.concat_after = concat_after
+        if self.concat_after:
+            self.concat_linear1 = nn.Linear(size + size, size)
+            self.concat_linear2 = nn.Linear(size + size, size)
+
+    def forward(self, tgt, tgt_mask, memory, memory_mask, cache=None):
+        """Compute decoded features.
+
+        Args:
+            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
+            tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
+            memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
+            memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
+            cache (List[torch.Tensor]): List of cached tensors.
+                Each tensor shape should be (#batch, maxlen_out - 1, size).
+
+        Returns:
+            torch.Tensor: Output tensor(#batch, maxlen_out, size).
+            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
+            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
+            torch.Tensor: Encoded memory mask (#batch, maxlen_in).
+
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        if cache is None:
+            tgt_q = tgt
+            tgt_q_mask = tgt_mask
+        else:
+            # compute only the last frame query keeping dim: max_time_out -> 1
+            assert cache.shape == (
+                tgt.shape[0],
+                tgt.shape[1] - 1,
+                self.size,
+            ), f"{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
+            tgt_q = tgt[:, -1:, :]
+            residual = residual[:, -1:, :]
+            tgt_q_mask = None
+            if tgt_mask is not None:
+                tgt_q_mask = tgt_mask[:, -1:, :]
+
+        if self.concat_after:
+            tgt_concat = torch.cat((tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)), dim=-1)
+            x = residual + self.concat_linear1(tgt_concat)
+        else:
+            x = residual + self.dropout(self.self_attn(tgt_q, tgt, tgt, tgt_q_mask))
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        if self.concat_after:
+            x_concat = torch.cat((x, self.src_attn(x, memory, memory, memory_mask)), dim=-1)
+            x = residual + self.concat_linear2(x_concat)
+        else:
+            x, score = self.src_attn(x, memory, memory, memory_mask)
+            x = residual + self.dropout(x)
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm3(x)
+        x = residual + self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm3(x)
+
+        if cache is not None:
+            x = torch.cat([cache, x], dim=1)
+
+        return x, tgt_mask, memory, memory_mask
+
+
+@tables.register("encoder_classes", "ConvBiasPredictor")
+class ConvPredictor(nn.Module):
+    def __init__(
+        self,
+        size=256,
+        l_order=3,
+        r_order=3,
+        attention_heads=4,
+        attention_dropout_rate=0.1,
+        linear_units=2048,
+    ):
+        """Initialize ConvPredictor.
+        
+            Args:
+                size: TODO.
+                l_order: TODO.
+                r_order: TODO.
+                attention_heads: TODO.
+                attention_dropout_rate: TODO.
+                linear_units: TODO.
+            """
+        super().__init__()
+        self.atten = MultiHeadedAttention(attention_heads, size, attention_dropout_rate)
+        self.norm1 = LayerNorm(size)
+        self.feed_forward = PositionwiseFeedForward(size, linear_units, attention_dropout_rate)
+        self.norm2 = LayerNorm(size)
+        self.pad = nn.ConstantPad1d((l_order, r_order), 0)
+        self.conv1d = nn.Conv1d(size, size, l_order + r_order + 1, groups=size)
+        self.output_linear = nn.Linear(size, 1)
+
+    def forward(self, text_enc, asr_enc):
+        # stage1 cross-attention
+        """Forward pass for training.
+        
+            Args:
+                text_enc: TODO.
+                asr_enc: TODO.
+            """
+        residual = text_enc
+        text_enc = residual + self.atten(text_enc, asr_enc, asr_enc, None)
+
+        # stage2 FFN
+        residual = text_enc
+        text_enc = self.norm1(text_enc)
+        text_enc = residual + self.feed_forward(text_enc)
+
+        # stage Conv predictor
+        text_enc = self.norm2(text_enc)
+        context = text_enc.transpose(1, 2)
+        queries = self.pad(context)
+        memory = self.conv1d(queries)
+        output = memory + context
+        output = output.transpose(1, 2)
+        output = torch.relu(output)
+        output = self.output_linear(output)
+        if output.dim() == 3:
+            output = output.squeeze(2)
+        return output

funasr/models/lcbnet/model.py

@@ -0,0 +1,574 @@
+#!/usr/bin/env python3
+# -*- encoding: utf-8 -*-
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+#  MIT License  (https://opensource.org/licenses/MIT)
+
+import logging
+from typing import Union, Dict, List, Tuple, Optional
+
+import time
+import torch
+import torch.nn as nn
+from torch.cuda.amp import autocast
+
+from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
+from funasr.models.ctc.ctc import CTC
+from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
+from funasr.metrics.compute_acc import th_accuracy
+
+# from funasr.models.e2e_asr_common import ErrorCalculator
+from funasr.train_utils.device_funcs import force_gatherable
+from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
+from funasr.utils import postprocess_utils
+from funasr.utils.datadir_writer import DatadirWriter
+from funasr.register import tables
+
+
+@tables.register("model_classes", "LCBNet")
+class LCBNet(nn.Module):
+    """LCBNet: Lightweight Convolutional Block Network for ASR.
+
+    Efficient model design using depthwise separable convolutions
+    for low-resource deployment scenarios.
+
+    Inherits Paraformer pipeline.
+    """
+
+    def __init__(
+        self,
+        specaug: str = None,
+        specaug_conf: dict = None,
+        normalize: str = None,
+        normalize_conf: dict = None,
+        encoder: str = None,
+        encoder_conf: dict = None,
+        decoder: str = None,
+        decoder_conf: dict = None,
+        text_encoder: str = None,
+        text_encoder_conf: dict = None,
+        bias_predictor: str = None,
+        bias_predictor_conf: dict = None,
+        fusion_encoder: str = None,
+        fusion_encoder_conf: dict = None,
+        ctc: str = None,
+        ctc_conf: dict = None,
+        ctc_weight: float = 0.5,
+        interctc_weight: float = 0.0,
+        select_num: int = 2,
+        select_length: int = 3,
+        insert_blank: bool = True,
+        input_size: int = 80,
+        vocab_size: int = -1,
+        ignore_id: int = -1,
+        blank_id: int = 0,
+        sos: int = 1,
+        eos: int = 2,
+        lsm_weight: float = 0.0,
+        length_normalized_loss: bool = False,
+        report_cer: bool = True,
+        report_wer: bool = True,
+        sym_space: str = "<space>",
+        sym_blank: str = "<blank>",
+        # extract_feats_in_collect_stats: bool = True,
+        share_embedding: bool = False,
+        # preencoder: Optional[AbsPreEncoder] = None,
+        # postencoder: Optional[AbsPostEncoder] = None,
+        **kwargs,
+    ):
+
+        """Initialize LCBNet.
+        
+            Args:
+                specaug: TODO.
+                specaug_conf: Configuration dict for specaug.
+                normalize: TODO.
+                normalize_conf: Configuration dict for normalize.
+                encoder: TODO.
+                encoder_conf: Configuration dict for encoder.
+                decoder: TODO.
+                decoder_conf: Configuration dict for decoder.
+                text_encoder: TODO.
+                text_encoder_conf: Configuration dict for text_encoder.
+                bias_predictor: TODO.
+                bias_predictor_conf: Configuration dict for bias_predictor.
+                fusion_encoder: TODO.
+                fusion_encoder_conf: Configuration dict for fusion_encoder.
+                ctc: TODO.
+                ctc_conf: Configuration dict for ctc.
+                ctc_weight: TODO.
+                interctc_weight: TODO.
+                select_num: TODO.
+                select_length: TODO.
+                insert_blank: TODO.
+                input_size: Size/dimension parameter.
+                vocab_size: Size/dimension parameter.
+                ignore_id: TODO.
+                blank_id: TODO.
+                sos: TODO.
+                eos: TODO.
+                lsm_weight: TODO.
+                length_normalized_loss: TODO.
+                report_cer: TODO.
+                report_wer: TODO.
+                sym_space: TODO.
+                sym_blank: TODO.
+                share_embedding: TODO.
+                **kwargs: Additional keyword arguments.
+            """
+        super().__init__()
+
+        if specaug is not None:
+            specaug_class = tables.specaug_classes.get(specaug)
+            specaug = specaug_class(**specaug_conf)
+        if normalize is not None:
+            normalize_class = tables.normalize_classes.get(normalize)
+            normalize = normalize_class(**normalize_conf)
+        encoder_class = tables.encoder_classes.get(encoder)
+        encoder = encoder_class(input_size=input_size, **encoder_conf)
+        encoder_output_size = encoder.output_size()
+
+        # lcbnet modules: text encoder, fusion encoder and bias predictor
+        text_encoder_class = tables.encoder_classes.get(text_encoder)
+        text_encoder = text_encoder_class(input_size=vocab_size, **text_encoder_conf)
+        fusion_encoder_class = tables.encoder_classes.get(fusion_encoder)
+        fusion_encoder = fusion_encoder_class(**fusion_encoder_conf)
+        bias_predictor_class = tables.encoder_classes.get(bias_predictor)
+        bias_predictor = bias_predictor_class(**bias_predictor_conf)
+
+        if decoder is not None:
+            decoder_class = tables.decoder_classes.get(decoder)
+            decoder = decoder_class(
+                vocab_size=vocab_size,
+                encoder_output_size=encoder_output_size,
+                **decoder_conf,
+            )
+        if ctc_weight > 0.0:
+
+            if ctc_conf is None:
+                ctc_conf = {}
+
+            ctc = CTC(odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf)
+
+        self.blank_id = blank_id
+        self.sos = vocab_size - 1
+        self.eos = vocab_size - 1
+        self.vocab_size = vocab_size
+        self.ignore_id = ignore_id
+        self.ctc_weight = ctc_weight
+        self.specaug = specaug
+        self.normalize = normalize
+        self.encoder = encoder
+        # lcbnet
+        self.text_encoder = text_encoder
+        self.fusion_encoder = fusion_encoder
+        self.bias_predictor = bias_predictor
+        self.select_num = select_num
+        self.select_length = select_length
+        self.insert_blank = insert_blank
+
+        if not hasattr(self.encoder, "interctc_use_conditioning"):
+            self.encoder.interctc_use_conditioning = False
+        if self.encoder.interctc_use_conditioning:
+            self.encoder.conditioning_layer = torch.nn.Linear(
+                vocab_size, self.encoder.output_size()
+            )
+        self.interctc_weight = interctc_weight
+
+        # self.error_calculator = None
+        if ctc_weight == 1.0:
+            self.decoder = None
+        else:
+            self.decoder = decoder
+
+        self.criterion_att = LabelSmoothingLoss(
+            size=vocab_size,
+            padding_idx=ignore_id,
+            smoothing=lsm_weight,
+            normalize_length=length_normalized_loss,
+        )
+        #
+        # if report_cer or report_wer:
+        #     self.error_calculator = ErrorCalculator(
+        #         token_list, sym_space, sym_blank, report_cer, report_wer
+        #     )
+        #
+        self.error_calculator = None
+        if ctc_weight == 0.0:
+            self.ctc = None
+        else:
+            self.ctc = ctc
+
+        self.share_embedding = share_embedding
+        if self.share_embedding:
+            self.decoder.embed = None
+
+        self.length_normalized_loss = length_normalized_loss
+        self.beam_search = None
+
+    def forward(
+        self,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
+        """Encoder + Decoder + Calc loss
+        Args:
+                speech: (Batch, Length, ...)
+                speech_lengths: (Batch, )
+                text: (Batch, Length)
+                text_lengths: (Batch,)
+        """
+
+        if len(text_lengths.size()) > 1:
+            text_lengths = text_lengths[:, 0]
+        if len(speech_lengths.size()) > 1:
+            speech_lengths = speech_lengths[:, 0]
+
+        batch_size = speech.shape[0]
+
+        # 1. Encoder
+        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
+        intermediate_outs = None
+        if isinstance(encoder_out, tuple):
+            intermediate_outs = encoder_out[1]
+            encoder_out = encoder_out[0]
+
+        loss_att, acc_att, cer_att, wer_att = None, None, None, None
+        loss_ctc, cer_ctc = None, None
+        stats = dict()
+
+        # decoder: CTC branch
+        if self.ctc_weight != 0.0:
+            loss_ctc, cer_ctc = self._calc_ctc_loss(
+                encoder_out, encoder_out_lens, text, text_lengths
+            )
+
+            # Collect CTC branch stats
+            stats["loss_ctc"] = loss_ctc.detach() if loss_ctc is not None else None
+            stats["cer_ctc"] = cer_ctc
+
+        # Intermediate CTC (optional)
+        loss_interctc = 0.0
+        if self.interctc_weight != 0.0 and intermediate_outs is not None:
+            for layer_idx, intermediate_out in intermediate_outs:
+                # we assume intermediate_out has the same length & padding
+                # as those of encoder_out
+                loss_ic, cer_ic = self._calc_ctc_loss(
+                    intermediate_out, encoder_out_lens, text, text_lengths
+                )
+                loss_interctc = loss_interctc + loss_ic
+
+                # Collect Intermedaite CTC stats
+                stats["loss_interctc_layer{}".format(layer_idx)] = (
+                    loss_ic.detach() if loss_ic is not None else None
+                )
+                stats["cer_interctc_layer{}".format(layer_idx)] = cer_ic
+
+            loss_interctc = loss_interctc / len(intermediate_outs)
+
+            # calculate whole encoder loss
+            loss_ctc = (1 - self.interctc_weight) * loss_ctc + self.interctc_weight * loss_interctc
+
+        # decoder: Attention decoder branch
+        loss_att, acc_att, cer_att, wer_att = self._calc_att_loss(
+            encoder_out, encoder_out_lens, text, text_lengths
+        )
+
+        # 3. CTC-Att loss definition
+        if self.ctc_weight == 0.0:
+            loss = loss_att
+        elif self.ctc_weight == 1.0:
+            loss = loss_ctc
+        else:
+            loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att
+
+        # Collect Attn branch stats
+        stats["loss_att"] = loss_att.detach() if loss_att is not None else None
+        stats["acc"] = acc_att
+        stats["cer"] = cer_att
+        stats["wer"] = wer_att
+
+        # Collect total loss stats
+        stats["loss"] = torch.clone(loss.detach())
+
+        # force_gatherable: to-device and to-tensor if scalar for DataParallel
+        if self.length_normalized_loss:
+            batch_size = int((text_lengths + 1).sum())
+        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
+        return loss, stats, weight
+
+    def encode(
+        self,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Frontend + Encoder. Note that this method is used by asr_inference.py
+        Args:
+                speech: (Batch, Length, ...)
+                speech_lengths: (Batch, )
+                ind: int
+        """
+        with autocast(False):
+            # Data augmentation
+            if self.specaug is not None and self.training:
+                speech, speech_lengths = self.specaug(speech, speech_lengths)
+            # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
+            if self.normalize is not None:
+                speech, speech_lengths = self.normalize(speech, speech_lengths)
+        # Forward encoder
+        # feats: (Batch, Length, Dim)
+        # -> encoder_out: (Batch, Length2, Dim2)
+        if self.encoder.interctc_use_conditioning:
+            encoder_out, encoder_out_lens, _ = self.encoder(speech, speech_lengths, ctc=self.ctc)
+        else:
+            encoder_out, encoder_out_lens, _ = self.encoder(speech, speech_lengths)
+        intermediate_outs = None
+        if isinstance(encoder_out, tuple):
+            intermediate_outs = encoder_out[1]
+            encoder_out = encoder_out[0]
+
+        if intermediate_outs is not None:
+            return (encoder_out, intermediate_outs), encoder_out_lens
+        return encoder_out, encoder_out_lens
+
+    def _calc_att_loss(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        ys_pad: torch.Tensor,
+        ys_pad_lens: torch.Tensor,
+    ):
+        """Internal: calc att loss.
+        
+            Args:
+                encoder_out: Encoder output tensor.
+                encoder_out_lens: Encoder output lengths.
+                ys_pad: TODO.
+                ys_pad_lens: Lengths of ys_pad.
+            """
+        ys_in_pad, ys_out_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
+        ys_in_lens = ys_pad_lens + 1
+
+        # 1. Forward decoder
+        decoder_out, _ = self.decoder(encoder_out, encoder_out_lens, ys_in_pad, ys_in_lens)
+
+        # 2. Compute attention loss
+        loss_att = self.criterion_att(decoder_out, ys_out_pad)
+        acc_att = th_accuracy(
+            decoder_out.view(-1, self.vocab_size),
+            ys_out_pad,
+            ignore_label=self.ignore_id,
+        )
+
+        # Compute cer/wer using attention-decoder
+        if self.training or self.error_calculator is None:
+            cer_att, wer_att = None, None
+        else:
+            ys_hat = decoder_out.argmax(dim=-1)
+            cer_att, wer_att = self.error_calculator(ys_hat.cpu(), ys_pad.cpu())
+
+        return loss_att, acc_att, cer_att, wer_att
+
+    def _calc_ctc_loss(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        ys_pad: torch.Tensor,
+        ys_pad_lens: torch.Tensor,
+    ):
+        # Calc CTC loss
+        """Internal: calc ctc loss.
+        
+            Args:
+                encoder_out: Encoder output tensor.
+                encoder_out_lens: Encoder output lengths.
+                ys_pad: TODO.
+                ys_pad_lens: Lengths of ys_pad.
+            """
+        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
+
+        # Calc CER using CTC
+        cer_ctc = None
+        if not self.training and self.error_calculator is not None:
+            ys_hat = self.ctc.argmax(encoder_out).data
+            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
+        return loss_ctc, cer_ctc
+
+    def init_beam_search(
+        self,
+        **kwargs,
+    ):
+        """Init beam search.
+        
+            Args:
+                **kwargs: Additional keyword arguments.
+            """
+        from funasr.models.transformer.search import BeamSearch
+        from funasr.models.transformer.scorers.ctc import CTCPrefixScorer
+        from funasr.models.transformer.scorers.length_bonus import LengthBonus
+
+        # 1. Build ASR model
+        scorers = {}
+
+        if self.ctc != None:
+            ctc = CTCPrefixScorer(ctc=self.ctc, eos=self.eos)
+            scorers.update(ctc=ctc)
+        token_list = kwargs.get("token_list")
+        scorers.update(
+            decoder=self.decoder,
+            length_bonus=LengthBonus(len(token_list)),
+        )
+
+        # 3. Build ngram model
+        # ngram is not supported now
+        ngram = None
+        scorers["ngram"] = ngram
+
+        weights = dict(
+            decoder=1.0 - kwargs.get("decoding_ctc_weight", 0.3),
+            ctc=kwargs.get("decoding_ctc_weight", 0.3),
+            lm=kwargs.get("lm_weight", 0.0),
+            ngram=kwargs.get("ngram_weight", 0.0),
+            length_bonus=kwargs.get("penalty", 0.0),
+        )
+        beam_search = BeamSearch(
+            beam_size=kwargs.get("beam_size", 20),
+            weights=weights,
+            scorers=scorers,
+            sos=self.sos,
+            eos=self.eos,
+            vocab_size=len(token_list),
+            token_list=token_list,
+            pre_beam_score_key=None if self.ctc_weight == 1.0 else "full",
+        )
+
+        self.beam_search = beam_search
+
+    def inference(
+        self,
+        data_in,
+        data_lengths=None,
+        key: list = None,
+        tokenizer=None,
+        frontend=None,
+        **kwargs,
+    ):
+
+        """Run inference on input data.
+        
+            Args:
+                data_in: Input data (audio samples, file paths, or text).
+                data_lengths: Lengths of each input sample in the batch.
+                key: Sample identifiers.
+                tokenizer: Tokenizer instance for text encoding/decoding.
+                frontend: Audio frontend for feature extraction.
+                **kwargs: Additional keyword arguments.
+            """
+        if kwargs.get("batch_size", 1) > 1:
+            raise NotImplementedError("batch decoding is not implemented")
+
+        # init beamsearch
+        if self.beam_search is None:
+            logging.info("enable beam_search")
+            self.init_beam_search(**kwargs)
+            self.nbest = kwargs.get("nbest", 1)
+
+        meta_data = {}
+        if (
+            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
+        ):  # fbank
+            speech, speech_lengths = data_in, data_lengths
+            if len(speech.shape) < 3:
+                speech = speech[None, :, :]
+            if speech_lengths is None:
+                speech_lengths = speech.shape[1]
+        else:
+            # extract fbank feats
+            time1 = time.perf_counter()
+            sample_list = load_audio_text_image_video(
+                data_in,
+                fs=frontend.fs,
+                audio_fs=kwargs.get("fs", 16000),
+                data_type=kwargs.get("data_type", "sound"),
+                tokenizer=tokenizer,
+            )
+            time2 = time.perf_counter()
+            meta_data["load_data"] = f"{time2 - time1:0.3f}"
+            audio_sample_list = sample_list[0]
+            if len(sample_list) > 1:
+                ocr_sample_list = sample_list[1]
+            else:
+                ocr_sample_list = [[294, 0]]
+            speech, speech_lengths = extract_fbank(
+                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
+            )
+            time3 = time.perf_counter()
+            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
+            frame_shift = 10
+            meta_data["batch_data_time"] = speech_lengths.sum().item() * frame_shift / 1000
+
+        speech = speech.to(device=kwargs["device"])
+        speech_lengths = speech_lengths.to(device=kwargs["device"])
+        # Encoder
+        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
+        if isinstance(encoder_out, tuple):
+            encoder_out = encoder_out[0]
+
+        ocr_list_new = [[x + 1 if x != 0 else x for x in sublist] for sublist in ocr_sample_list]
+        ocr = torch.tensor(ocr_list_new).to(device=kwargs["device"])
+        ocr_lengths = ocr.new_full([1], dtype=torch.long, fill_value=ocr.size(1)).to(
+            device=kwargs["device"]
+        )
+        ocr, ocr_lens, _ = self.text_encoder(ocr, ocr_lengths)
+        fusion_out, _, _, _ = self.fusion_encoder(encoder_out, None, ocr, None)
+        encoder_out = encoder_out + fusion_out
+        # c. Passed the encoder result and the beam search
+        nbest_hyps = self.beam_search(
+            x=encoder_out[0],
+            maxlenratio=kwargs.get("maxlenratio", 0.0),
+            minlenratio=kwargs.get("minlenratio", 0.0),
+        )
+
+        nbest_hyps = nbest_hyps[: self.nbest]
+
+        results = []
+        b, n, d = encoder_out.size()
+        for i in range(b):
+
+            for nbest_idx, hyp in enumerate(nbest_hyps):
+                ibest_writer = None
+                if kwargs.get("output_dir") is not None:
+                    if not hasattr(self, "writer"):
+                        self.writer = DatadirWriter(kwargs.get("output_dir"))
+                    ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]
+
+                # remove sos/eos and get results
+                last_pos = -1
+                if isinstance(hyp.yseq, list):
+                    token_int = hyp.yseq[1:last_pos]
+                else:
+                    token_int = hyp.yseq[1:last_pos].tolist()
+
+                # remove blank symbol id, which is assumed to be 0
+                token_int = list(
+                    filter(
+                        lambda x: x != self.eos and x != self.sos and x != self.blank_id, token_int
+                    )
+                )
+
+                # Change integer-ids to tokens
+                token = tokenizer.ids2tokens(token_int)
+                text = tokenizer.tokens2text(token)
+
+                text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
+                result_i = {"key": key[i], "token": token, "text": text_postprocessed}
+                results.append(result_i)
+
+                if ibest_writer is not None:
+                    ibest_writer["token"][key[i]] = " ".join(token)
+                    ibest_writer["text"][key[i]] = text_postprocessed
+
+        return results, meta_data