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
fixbug (#1412)
zhifu gao5608bee0acce
diff --git a/funasr/models/contextual_paraformer/model.py b/funasr/models/contextual_paraformer/model.pyindex 939d46d64..49868a8f4 100644--- a/funasr/models/contextual_paraformer/model.py+++ b/funasr/models/contextual_paraformer/model.py@@ -190,13 +190,10 @@ class ContextualParaformer(Paraformer):# 0. samplerdecoder_out_1st = Noneif self.sampling_ratio > 0.0:- if self.step_cur < 2:- logging.info("enable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))+sematic_embeds, decoder_out_1st = self.sampler(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens,pre_acoustic_embeds, contextual_info)else:- if self.step_cur < 2:- logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))sematic_embeds = pre_acoustic_embeds# 1. Forward decoderdiff --git a/funasr/models/paraformer/model.py b/funasr/models/paraformer/model.pyindex 90ce162d4..51c9bdb47 100644--- a/funasr/models/paraformer/model.py+++ b/funasr/models/paraformer/model.py@@ -154,8 +154,8 @@ class Paraformer(torch.nn.Module):self.predictor_bias = predictor_biasself.sampling_ratio = sampling_ratioself.criterion_pre = mae_loss(normalize_length=length_normalized_loss)- # self.step_cur = 0- #++self.share_embedding = share_embeddingif self.share_embedding:self.decoder.embed = Nonediff --git a/funasr/models/paraformer_streaming/model.py b/funasr/models/paraformer_streaming/model.pyindex 45875a2da..4cf20dec5 100644--- a/funasr/models/paraformer_streaming/model.py+++ b/funasr/models/paraformer_streaming/model.py@@ -235,8 +235,7 @@ class ParaformerStreaming(Paraformer):decoder_out_1st = Nonepre_loss_att = Noneif self.sampling_ratio > 0.0:- if self.step_cur < 2:- logging.info("enable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))+if self.use_1st_decoder_loss:sematic_embeds, decoder_out_1st, pre_loss_att = \self.sampler_with_grad(encoder_out, encoder_out_lens, ys_pad,@@ -246,8 +245,6 @@ class ParaformerStreaming(Paraformer):self.sampler(encoder_out, encoder_out_lens, ys_pad,ys_pad_lens, pre_acoustic_embeds, scama_mask)else:- if self.step_cur < 2:- logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))sematic_embeds = pre_acoustic_embeds# 1. Forward decoderdiff --git a/funasr/models/seaco_paraformer/model.py b/funasr/models/seaco_paraformer/model.pyindex 21ad87499..20b0cc838 100644--- a/funasr/models/seaco_paraformer/model.py+++ b/funasr/models/seaco_paraformer/model.py@@ -130,7 +130,6 @@ class SeacoParaformer(BiCifParaformer, Paraformer):dha_pad = kwargs.get("dha_pad")batch_size = speech.shape[0]- self.step_cur += 1# for data-paralleltext = text[:, : text_lengths.max()]speech = speech[:, :speech_lengths.max()]