Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
auto frontend
游雁162efb747f21
diff --git a/funasr/models/llm_asr/adaptor.py b/funasr/models/llm_asr/adaptor.pyindex 9b79ed28a..c93988328 100644--- a/funasr/models/llm_asr/adaptor.py+++ b/funasr/models/llm_asr/adaptor.py@@ -83,25 +83,27 @@ class Transformer(nn.Module):from funasr.models.transformer.attention import MultiHeadedAttentionfrom funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward- self.blocks = nn.ModuleList(- [- EncoderLayer(- llm_dim,- MultiHeadedAttention(- kwargs.get("attention_heads", 8),+ self.blocks = None+ if kwargs.get("n_layer", 2) > 0:+ self.blocks = nn.ModuleList(+ [+ EncoderLayer(llm_dim,- kwargs.get("attention_dropout_rate", 0.0),- ),- PositionwiseFeedForward(- llm_dim,- llm_dim // 4,+ MultiHeadedAttention(+ kwargs.get("attention_heads", 8),+ llm_dim,+ kwargs.get("attention_dropout_rate", 0.0),+ ),+ PositionwiseFeedForward(+ llm_dim,+ llm_dim // 4,+ kwargs.get("dropout_rate", 0.0),+ ),kwargs.get("dropout_rate", 0.0),- ),- kwargs.get("dropout_rate", 0.0),- )- for i in range(kwargs.get("n_layer", 2))- ]- )+ )+ for i in range(kwargs.get("n_layer", 2))+ ]+ )def forward(self, x, ilens=None):@@ -123,6 +125,7 @@ class Transformer(nn.Module):olens = Noneolens = (ilens - 1) // self.k + 1masks = (~make_pad_mask(olens)[:, None, :]).to(x.device)- for layer, block in enumerate(self.blocks):- x, masks = block(x, masks)+ if self.blocks is not None:+ for layer, block in enumerate(self.blocks):+ x, masks = block(x, masks)return x, olensdiff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex b13912346..d94058cca 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -481,7 +481,7 @@ class LLMASR2(nn.Module):batch_size, token_num, dims = inputs_embeds.shapefbank_mask[fbank_mask < 0] = 0- fbank_fake_lens = fbank_mask.sum(-1)+ fbank_fake_lens = fbank_mask.sum(-1).to(torch.int32)# _, l, _ = encoder_out.shapefor batch_idx in range(batch_size):