app.py

from svoice.separate import *
import scipy.io.wavfile as wav
import gradio as gr
import os
import torch
import soundfile as sf
from transformers import pipeline
from glob import glob
load_model()

device = "cuda" if torch.cuda.is_available() else "cpu"
BASE_PATH = os.path.dirname(os.path.abspath(__file__))
os.makedirs('input', exist_ok=True)
os.makedirs('separated', exist_ok=True)

print("Loading ASR model...")
pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=0 if device == "cuda" else -1)
print("ASR model loaded!")

def transcribe_audio(audiopath):
    audio_input, sr = sf.read(audiopath)
    return pipe(audio_input, sampling_rate=sr, return_tensors=False, padding=True, max_new_tokens=500)['text']

def separator(audio, rec_audio, example):
    outputs= {}
    for f in glob('input/*'):
        os.remove(f)
    for f in glob('separated/*'):
        os.remove(f)
    if audio:
        wav.write('input/original.wav', audio[0], audio[1])
    elif rec_audio:
        wav.write('input/original.wav', rec_audio[0], rec_audio[1])
    else:
        os.system(f'cp {example} input/original.wav')
    separate_demo(mix_dir="./input")
    separated_files = glob(os.path.join('separated', "*.wav"))
    separated_files = sorted([f for f in separated_files if "original.wav" not in f])
    outputs["transcripts"] = []
    
    for i, f in enumerate(separated_files):
        print(f"Transcribing separated audio {i+1} ...")
        outputs["transcripts"].append(transcribe_audio(f))
        print("Text:", outputs["transcripts"][-1])
    return separated_files + outputs['transcripts']

def set_example_audio(example: list) -> dict:
    return gr.Audio.update(value=example[0])

demo = gr.Blocks()
with demo:
    gr.Markdown('''
    <center>
        <h1>Multiple Voice Separation with Transcription DEMO</h1>
        <div style="display:flex;align-items:center;justify-content:center;"><iframe src="https://streamable.com/e/0x8osl?autoplay=1&nocontrols=1" frameborder="0" allow="autoplay"></iframe></div>
        <p>
            This is a demo for the multiple voice separation algorithm. The algorithm is trained on the LibriMix7 dataset and can be used to separate multiple voices from a single audio file.
        </p>
    </center>
    ''')
    
    with gr.Row():
        input_audio = gr.Audio(label="Input audio", type="numpy")
        rec_audio = gr.Audio(label="Record Using Microphone", type="numpy", source="microphone")

    with gr.Row():
        output_audio1 = gr.Audio(label='Speaker 1', interactive=False)
        output_text1 = gr.Text(label='Speaker 1', interactive=False)
        output_audio2 = gr.Audio(label='Speaker 2', interactive=False)
        output_text2 = gr.Text(label='Speaker 2', interactive=False)

    with gr.Row():
        output_audio3 = gr.Audio(label='Speaker 3', interactive=False)
        output_text3 = gr.Text(label='Speaker 3', interactive=False)
        output_audio4 = gr.Audio(label='Speaker 4', interactive=False)
        output_text4 = gr.Text(label='Speaker 4', interactive=False)

    with gr.Row():
        output_audio5 = gr.Audio(label='Speaker 5', interactive=False)
        output_text5 = gr.Text(label='Speaker 5', interactive=False)
        output_audio6 = gr.Audio(label='Speaker 6', interactive=False)
        output_text6 = gr.Text(label='Speaker 6', interactive=False)

    with gr.Row():
        output_audio7 = gr.Audio(label='Speaker 7', interactive=False)
        output_text7 = gr.Text(label='Speaker 7', interactive=False)

    outputs_audio = [output_audio1, output_audio2, output_audio3, output_audio4, output_audio5, output_audio6, output_audio7]
    outputs_text = [output_text1, output_text2, output_text3, output_text4, output_text5, output_text6, output_text7]
    button = gr.Button("Separate")
    examples = [
        "samples/mixture1.wav",
        "samples/mixture2.wav",
        "samples/mixture3.wav"
    ]
    example_selector = gr.inputs.Radio(examples, label="Example Audio")
    button.click(separator, inputs=[input_audio, rec_audio, example_selector], outputs=outputs_audio + outputs_text)

demo.launch()