提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py

@@ -62,7 +62,7 @@
        if quantize:
            model_file = os.path.join(model_dir, "model_quant.onnx")
        if not os.path.exists(model_file):
            print(".onnx is not exist, begin to export onnx")
            print(".onnx does not exist, begin to export onnx")
            try:
                from funasr import AutoModel
            except:
@@ -175,7 +175,23 @@
        plt.savefig(plotname, bbox_inches="tight")

    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

@@ -285,7 +301,7 @@
            model_eb_file = os.path.join(model_dir, "model_eb.onnx")

        if not (os.path.exists(model_eb_file) and os.path.exists(model_bb_file)):
            print(".onnx is not exist, begin to export onnx")
            print(".onnx does not exist, begin to export onnx")
            try:
                from funasr import AutoModel
            except:
@@ -322,6 +338,10 @@
            self.pred_bias = config["model_conf"]["predictor_bias"]
        else:
            self.pred_bias = 0
        if "lang" in config:
            self.language = config["lang"]
        else:
            self.language = None

    def __call__(
        self, wav_content: Union[str, np.ndarray, List[str]], hotwords: str, **kwargs
@@ -331,7 +351,6 @@
    # ) -> List:
        # make hotword list
        hotwords, hotwords_length = self.proc_hotword(hotwords)
        # import pdb; pdb.set_trace()
        [bias_embed] = self.eb_infer(hotwords, hotwords_length)
        # index from bias_embed
        bias_embed = bias_embed.transpose(1, 0, 2)
@@ -411,10 +430,10 @@
            return np.array(hotwords)

        hotword_int = [word_map(i) for i in hotwords]
        # import pdb; pdb.set_trace()

        hotword_int.append(np.array([1]))
        hotwords = pad_list(hotword_int, pad_value=0, max_len=10)
        # import pdb; pdb.set_trace()
        
        return hotwords, hotwords_length

    def bb_infer(