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
update seaco finetune (#1526)
Shi Xian980007a486a4
diff --git a/examples/industrial_data_pretraining/seaco_paraformer/finetune.sh b/examples/industrial_data_pretraining/seaco_paraformer/finetune.shindex cfdec7779..5614f44a5 100644--- a/examples/industrial_data_pretraining/seaco_paraformer/finetune.sh+++ b/examples/industrial_data_pretraining/seaco_paraformer/finetune.sh@@ -10,7 +10,7 @@ gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')## option 1, download model automaticallymodel_name_or_model_dir="iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"-model_revision="v2.0.4"+model_revision="v2.0.7"## option 2, download model by git#local_path_root=${workspace}/modelscope_modelsdiff --git a/funasr/models/contextual_paraformer/model.py b/funasr/models/contextual_paraformer/model.pyindex 9968bf2e4..b9fd3c463 100644--- a/funasr/models/contextual_paraformer/model.py+++ b/funasr/models/contextual_paraformer/model.py@@ -94,10 +94,8 @@ class ContextualParaformer(Paraformer):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]+ text_lengths = text_lengths.squeeze()+ speech_lengths = speech_lengths.squeeze()batch_size = speech.shape[0]diff --git a/funasr/models/seaco_paraformer/model.py b/funasr/models/seaco_paraformer/model.pyindex 27ff5d1f8..21b6abaec 100644--- a/funasr/models/seaco_paraformer/model.py+++ b/funasr/models/seaco_paraformer/model.py@@ -117,6 +117,8 @@ class SeacoParaformer(BiCifParaformer, Paraformer):text: (Batch, Length)text_lengths: (Batch,)"""+ text_lengths = text_lengths.squeeze()+ speech_lengths = speech_lengths.squeeze()assert text_lengths.dim() == 1, text_lengths.shape# Check that batch_size is unifiedassert (@@ -164,7 +166,7 @@ class SeacoParaformer(BiCifParaformer, Paraformer):# force_gatherable: to-device and to-tensor if scalar for DataParallelif self.length_normalized_loss:- batch_size = (text_lengths + self.predictor_bias).sum().type_as(batch_size)+ batch_size = (text_lengths + self.predictor_bias).sum()loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)return loss, stats, weight@@ -190,8 +192,7 @@ class SeacoParaformer(BiCifParaformer, Paraformer):# predictor forwardencoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(encoder_out.device)- pre_acoustic_embeds, _, _, _ = self.predictor(encoder_out, ys_pad, encoder_out_mask,- ignore_id=self.ignore_id)+ pre_acoustic_embeds = self.predictor(encoder_out, ys_pad, encoder_out_mask, ignore_id=self.ignore_id)[0]# decoder forwarddecoder_out, _ = self.decoder(encoder_out, encoder_out_lens, pre_acoustic_embeds, ys_lengths, return_hidden=True)selected = self._hotword_representation(hotword_pad,