提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py

@@ -3,8 +3,6 @@
# Copyright FunASR (https://github.com/FunAudioLLM/SenseVoice). All Rights Reserved.
#  MIT License  (https://opensource.org/licenses/MIT)


import torch
import os.path
import librosa
import numpy as np
@@ -94,7 +92,7 @@
            return self.lid_dict[lid]
        else:
            raise ValueError(
                f"The language {l} is not in {list(self.lid_dict.keys())}"
                f"The language {lid} is not in {list(self.lid_dict.keys())}"
            )
            
    def _get_tnid(self, tnid):
@@ -181,12 +179,14 @@
            )
            for b in range(feats.shape[0]):
                # back to torch.Tensor
                if isinstance(ctc_logits, np.ndarray):
                    ctc_logits = torch.from_numpy(ctc_logits).float()
                # if isinstance(ctc_logits, np.ndarray):
                #     ctc_logits = torch.from_numpy(ctc_logits).float()
                # support batch_size=1 only currently
                x = ctc_logits[b, : encoder_out_lens[b].item(), :]
                yseq = x.argmax(dim=-1)
                yseq = torch.unique_consecutive(yseq, dim=-1)
                yseq = np.argmax(x, axis=-1)
                # Use np.diff and np.where instead of torch.unique_consecutive.
                mask = np.concatenate(([True], np.diff(yseq) != 0))
                yseq = yseq[mask]

                mask = yseq != self.blank_id
                token_int = yseq[mask].tolist()
@@ -196,7 +196,24 @@
        return asr_res

    def load_data(self, wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List:
        
        def convert_to_wav(input_path, output_path):
            from pydub import AudioSegment
            try:
                audio = AudioSegment.from_mp3(input_path)
                audio.export(output_path, format="wav")
                print("音频文件为mp3格式，已转换为wav格式")
                
            except Exception as e:
                print(f"转换失败:{e}")

        def load_wav(path: str) -> np.ndarray:
            if not path.lower().endswith('.wav'):
                import os
                input_path = path
                path = os.path.splitext(path)[0]+'.wav'
                convert_to_wav(input_path,path) #将mp3格式转换成wav格式

            waveform, _ = librosa.load(path, sr=fs)
            return waveform

@@ -215,6 +232,10 @@
        feats, feats_len = [], []
        for waveform in waveform_list:
            speech, _ = self.frontend.fbank(waveform)

            if speech is None or speech.size == 0:
                print("detected speech size {speech.size}")
                raise ValueError("Empty speech detected, skipping this waveform.")
            feat, feat_len = self.frontend.lfr_cmvn(speech)
            feats.append(feat)
            feats_len.append(feat_len)