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
EMO_UNK禁用和Merge VAD修复 (#1940)
gaochangfeng340c55838b56
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex d0a77337f..acd4c7229 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -337,7 +337,7 @@ class AutoModel:end_vad = time.time()# FIX(gcf): concat the vad clips for sense vocie model for better aed- if kwargs.get("merge_vad", False):+ if cfg.get("merge_vad", False):for i in range(len(res)):res[i]["value"] = merge_vad(res[i]["value"], kwargs.get("merge_length_s", 15) * 1000diff --git a/funasr/models/sense_voice/model.py b/funasr/models/sense_voice/model.pyindex cf4f7fb6a..5159b82b0 100644--- a/funasr/models/sense_voice/model.py+++ b/funasr/models/sense_voice/model.py@@ -644,6 +644,7 @@ class SenseVoiceSmall(nn.Module):self.embed = torch.nn.Embedding(7 + len(self.lid_dict) + len(self.textnorm_dict), input_size)+ self.emo_dict = {"unk": 25009, "happy": 25001, "sad": 25002, "angry": 25003, "neutral": 25004}self.criterion_att = LabelSmoothingLoss(size=self.vocab_size,@@ -870,7 +871,9 @@ class SenseVoiceSmall(nn.Module):# c. Passed the encoder result and the beam searchctc_logits = self.ctc.log_softmax(encoder_out)-+ if kwargs.get("ban_emo_unk", False):+ ctc_logits[:, :, self.emo_dict["unk"]] = -float("inf")+results = []b, n, d = encoder_out.size()if isinstance(key[0], (list, tuple)):diff --git a/funasr/utils/vad_utils.py b/funasr/utils/vad_utils.pyindex eba48a972..f87a85e71 100644--- a/funasr/utils/vad_utils.py+++ b/funasr/utils/vad_utils.py@@ -32,8 +32,10 @@ def slice_padding_audio_samples(speech, speech_lengths, vad_segments):return speech_list, speech_lengths_list-def merge_vad(vad_result, max_length=15000):+def merge_vad(vad_result, max_length=15000, min_length=0):new_result = []+ if len(vad_result) <= 1:+ return vad_resulttime_step = [t[0] for t in vad_result] + [t[1] for t in vad_result]time_step = sorted(list(set(time_step)))if len(time_step) == 0:@@ -43,13 +45,15 @@ def merge_vad(vad_result, max_length=15000):time = time_step[i]if time_step[i + 1] - bg < max_length:continue- if time - bg < max_length * 1.5:+ if time - bg > min_length:new_result.append([bg, time])- else:- split_num = int(time - bg) // max_length + 1- spl_l = int(time - bg) // split_num- for j in range(split_num):- new_result.append([bg + j * spl_l, bg + (j + 1) * spl_l])+ # if time - bg < max_length * 1.5:+ # new_result.append([bg, time])+ # else:+ # split_num = int(time - bg) // max_length + 1+ # spl_l = int(time - bg) // split_num+ # for j in range(split_num):+ # new_result.append([bg + j * spl_l, bg + (j + 1) * spl_l])bg = timenew_result.append([bg, time_step[-1]])- return new_result+ return new_result\ No newline at end of file