Function separate_flowsep

app.py:226–252

(audio_path, text)

def separate_flowsep(audio_path, text):
    device = get_runtime_device()
    model, preprocessor = load_flowsep()
    full_wav = preprocessor.load_full_audio(audio_path)
    input_len = full_wav.shape[0]

    with torch.no_grad():
        if input_len <= FLOWSEP_CHUNK_IN:
            sep_audio = _flowsep_process_chunk(model, preprocessor, full_wav.copy(), text)
        else:
            out_list = []
            start = 0
            while start < input_len:
                end = min(start + FLOWSEP_CHUNK_IN, input_len)
                chunk = full_wav[start:end]
                out_chunk = _flowsep_process_chunk(model, preprocessor, chunk.copy(), text)
                need = min(FLOWSEP_CHUNK_OUT, input_len - start)
                out_list.append(out_chunk[:need])
                start += FLOWSEP_CHUNK_OUT
            sep_audio = np.concatenate(out_list)

        if len(sep_audio) > input_len:
            sep_audio = sep_audio[:input_len]
        elif len(sep_audio) < input_len:
            sep_audio = np.pad(sep_audio, (0, input_len - len(sep_audio)), mode="constant", constant_values=0)

    return (FLOWSEP_SR, sep_audio)


def inference(audio, text, model_choice):