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
add
游雁da340e6a6cf8
diff --git a/examples/industrial_data_pretraining/sense_voice/demo.py b/examples/industrial_data_pretraining/sense_voice/demo.pyindex 506b069dd..46ec93059 100644--- a/examples/industrial_data_pretraining/sense_voice/demo.py+++ b/examples/industrial_data_pretraining/sense_voice/demo.py@@ -9,8 +9,15 @@ model = AutoModel(model="/Users/zhifu/Downloads/modelscope_models/SenseVoice",vad_model="iic/speech_fsmn_vad_zh-cn-16k-common-pytorch",vad_kwargs={"max_single_segment_time": 30000},)-task = "ASR"-language = None-input_wav = "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"-res = model.generate(task=task, language=language, input=input_wav, batch_size_s=0,)+++input_wav = "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/SenseVoice/aed_ser/asr_bgm.wav"++DecodingOptions = {+ "task": ("ASR", "AED", "SER"),+ "language": "auto",+ "fp16": True,+ }++res = model.generate(input=input_wav, batch_size_s=0, DecodingOptions=DecodingOptions)print(res)diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex 830d88c94..630c390f7 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -211,6 +211,9 @@ class AutoModel:else:print(f"error, init_param does not exist!: {init_param}")+ # fp16+ if kwargs.get("fp16", False):+ model.to(torch.float16)return model, kwargsdef __call__(self, *args, **cfg):diff --git a/funasr/models/sense_voice/model.py b/funasr/models/sense_voice/model.pyindex 521dec888..f8c11778f 100644--- a/funasr/models/sense_voice/model.py+++ b/funasr/models/sense_voice/model.py@@ -73,28 +73,27 @@ class SenseVoice(nn.Module):speech = speech.to(device=kwargs["device"])[0, :, :]speech_lengths = speech_lengths.to(device=kwargs["device"])-- task = kwargs.get("task", "ASR")++ DecodingOptions = kwargs.get("DecodingOptions", {})+ task = DecodingOptions.get("task", "ASR")if isinstance(task, str):task = [task]task = "".join([f"<|{x}|>" for x in task])initial_prompt = kwargs.get("initial_prompt", f"<|startoftranscript|>{task}")- language = kwargs.get("language", None)- language = None if language == "auto" else language- # if language is None:- # # detect the spoken language- # _, probs = self.model.detect_language(speech, initial_prompt=initial_prompt)- # print(f"Detected language: {max(probs, key=probs.get)}")- # language = max(probs, key=probs.get)- # language = language if kwargs.get("language", None) is None else kwargs.get("language")+ DecodingOptions["initial_prompt"] = initial_prompt- # decode the audio+ language = DecodingOptions.get("language", None)+ language = None if language == "auto" else language+ DecodingOptions["language"] = language++ DecodingOptions["vocab_path"] = kwargs.get("vocab_path", None)- # initial_prompt = kwargs.get("initial_prompt", "<|startoftranscript|><|ASR|>")- vocab_path = kwargs.get("vocab_path", None)- options = whisper.DecodingOptions(language=language, fp16=False, without_timestamps=True, initial_prompt=initial_prompt, vocab_path=vocab_path)+ if "without_timestamps" not in DecodingOptions:+ DecodingOptions["without_timestamps"] = True++ options = whisper.DecodingOptions(**DecodingOptions)result = whisper.decode(self.model, speech, options)