app.py

import gradio as gr
import os
import soundfile as sf
import numpy as np


from decode import InferencePipeline
from datahandler import AudioMixer, fix_audio_format
from omegaconf import OmegaConf

MODEL_CACHE = {
    "base_model": InferencePipeline(OmegaConf.load("config/config.yaml")),
    "iter_model": InferencePipeline(OmegaConf.load("config/config_ira.yaml"))
}

# cfg = OmegaConf.load("config/config_ira.yaml")
# inter = InferencePipeline(cfg)
datamix = AudioMixer()


def gradio_TSE(input_audio_path, enroll_audio_path1, audio_type, model_select):
    print(f"模型选择: {model_select}")
    print(f"User uploaded audio path: {input_audio_path}")
    print(f"User enroll audio path:  {enroll_audio_path1}")
    # print(f"User enroll audio path:  {enroll_audio_path2}")

    # if model_select == "base_model":
    #     cfg_path = "config/config_base.yaml"
    # elif model_select == "iter_model":
    #     cfg_path = "config/config_iter.yaml"
    # else:
    #     raise ValueError("未知模型类型")
    
    inter = MODEL_CACHE[model_select]

    audio_info = sf.info(input_audio_path)
    print(f"采样率: {audio_info.samplerate} Hz")

    eol_info = sf.info(enroll_audio_path1)
    print(f"采样率: {eol_info.samplerate} Hz")

    # eol_info = sf.info(enroll_audio_path2)
    # print(f"采样率: {eol_info.samplerate} Hz")

    input_deal_wav = fix_audio_format(input_audio_path)
    input_deal_wav_path = "deal_input.wav"
    sf.write(input_deal_wav_path, input_deal_wav, 16000)

    enroll_wav1 = fix_audio_format(enroll_audio_path1)
    eol_wav1 = "eol1.wav"
    sf.write(eol_wav1, enroll_wav1, 16000)

    # if len(enroll_wav1) > 4 * 16000:
    #     middle_start = (len(enroll_wav1) - 3 * 16000) // 2
    #     middle_end = middle_start + 3 * 16000
    #     enroll_wav2 = enroll_wav1[middle_start:middle_end]
    #     print("成功提取 enroll_wav2，长度:", len(enroll_wav2))
    #     eol_wav2 = "eol2.wav"
    #     sf.write(eol_wav2, enroll_wav2, 16000)
    # else:
    # enroll_wav2 = fix_audio_format(enroll_audio_path2)
    # print("成功导入 enroll_wav2，长度:", len(enroll_wav2))
    # eol_wav2 = "eol2.wav"
    # sf.write(eol_wav2, enroll_wav2, 16000)

    if audio_type == "clean":
        noise_folder_test = "noises/" 

        mix_path = datamix.mix_with_noise_folder(input_deal_wav_path,noise_folder_test)
        print(f"Converted clean -> mix: {mix_path}")
    else:

        mix_path = input_deal_wav_path


    est_path1 = inter.run_inference(mix_path, eol_wav1)

    # est_path2 = inter.run_inference(mix_path, eol_wav2)

    return mix_path,est_path1



with gr.Blocks() as demo:
    gr.Markdown("## Target Speaker Extraction Demo")
    gr.Markdown(
        "This demo can handle either clean audio (which we'll turn into a mix) or directly a mix audio. Due to limited training data, recording directly with a mobile browser with music/vocal noises is more likely to achieve good results"
    )

    with gr.Row():
        with gr.Column(scale=2):
            input_audio = gr.Audio(label="Upload/record your audio", type="filepath")
        with gr.Column(scale=1):
            audio_type = gr.Radio(
                choices=["clean", "mix"],
                value="clean",
                label="Input audio type?"
            )
            model_select = gr.Radio(
                choices=["base_model", "iter_model"],
                value="iter_model",
                label="Select Model Type"
            )
    with gr.Row():

        enroll_audio1 = gr.Audio(label="Upload your first enroll audio", type="filepath")
        # enroll_audio2 = gr.Audio(label="Upload your second enroll audio to compare", type="filepath")
    
    with gr.Row():

        noisy_audio_output = gr.Audio(label="Noisy Audio (Processed input audio)", type="filepath")

    with gr.Row():
        extracted_audio_output1 = gr.Audio(label="First enroll extracted target speaker audio", type="filepath")
        # extracted_audio_output2 = gr.Audio(label="Second enroll extracted target speaker audio", type="filepath")

    convert_button = gr.Button("Extract")
    convert_button.click(
        fn=gradio_TSE,
        inputs=[input_audio, enroll_audio1, audio_type, model_select],
        outputs=[noisy_audio_output, extracted_audio_output1]
    )

if __name__ == "__main__":
    demo.launch()