Function tokenizer

utils/cosyvoice_tokenizer.py:22–27

(audio_segment)

def extract_speech_token(ort_session, wav_path, pool_executor=None):
    def tokenizer(audio_segment):
        feat = whisper.log_mel_spectrogram(audio_segment, n_mels=128)
        speech_token = ort_session.run(None, {ort_session.get_inputs()[0].name: feat.detach().cpu().numpy(),
                                              ort_session.get_inputs()[1].name: np.array([feat.shape[2]], dtype=np.int32)})[
            0].flatten().tolist()
        return speech_token

    if isinstance(wav_path, str):
        audio, sample_rate = torchaudio.load(wav_path, backend='soundfile')