Initial commit: FunASR Speech Recognition Toolkit

Add complete FunASR codebase including models, runtime, and documentation.
2026-07-09 22:38:58 +08:00
commit 6116b1f3c6
3683 changed files with 990984 additions and 0 deletions
@@ -0,0 +1,55 @@
+import torch
+from torch.nn import functional as F
+from funasr.models.decoder.abs_decoder import AbsDecoder
+
+
+class DenseDecoder(AbsDecoder):
+    def __init__(
+        self,
+        vocab_size,
+        encoder_output_size,
+        num_nodes_resnet1: int = 256,
+        num_nodes_last_layer: int = 256,
+        batchnorm_momentum: float = 0.5,
+    ):
+        """Initialize DenseDecoder.
+        
+            Args:
+                vocab_size: Size/dimension parameter.
+                encoder_output_size: Size/dimension parameter.
+                num_nodes_resnet1: TODO.
+                num_nodes_last_layer: TODO.
+                batchnorm_momentum: TODO.
+            """
+        super(DenseDecoder, self).__init__()
+        self.resnet1_dense = torch.nn.Linear(encoder_output_size, num_nodes_resnet1)
+        self.resnet1_bn = torch.nn.BatchNorm1d(
+            num_nodes_resnet1, eps=1e-3, momentum=batchnorm_momentum
+        )
+
+        self.resnet2_dense = torch.nn.Linear(num_nodes_resnet1, num_nodes_last_layer)
+        self.resnet2_bn = torch.nn.BatchNorm1d(
+            num_nodes_last_layer, eps=1e-3, momentum=batchnorm_momentum
+        )
+
+        self.output_dense = torch.nn.Linear(num_nodes_last_layer, vocab_size, bias=False)
+
+    def forward(self, features):
+        """Forward pass for training.
+        
+            Args:
+                features: TODO.
+            """
+        embeddings = {}
+        features = self.resnet1_dense(features)
+        embeddings["resnet1_dense"] = features
+        features = F.relu(features)
+        features = self.resnet1_bn(features)
+
+        features = self.resnet2_dense(features)
+        embeddings["resnet2_dense"] = features
+        features = F.relu(features)
+        features = self.resnet2_bn(features)
+
+        features = self.output_dense(features)
+        return features, embeddings