app.py

import gradio as gr
import torch
import numpy as np
import librosa

from torchmetrics.functional.audio.nisqa import non_intrusive_speech_quality_assessment as tm_nisqa

SR = 16000

def predict_nisqa(audio):
    if isinstance(audio, tuple):
        _sr, y = audio
        y = librosa.resample(np.asarray(y).astype(np.float32), orig_sr=_sr, target_sr=SR)
    else:
        y, _ = librosa.load(audio, sr=SR, mono=True)

    wav = torch.tensor(y, dtype=torch.float32)

    out = tm_nisqa(wav, SR).detach().cpu().numpy().tolist()
    mos, noisiness, discontinuity, coloration, loudness = out

    table = {
        "Metric": ["MOS (overall)", "Noisiness", "Discontinuity", "Coloration", "Loudness"],
        "Score":  [round(mos, 3), round(noisiness, 3), round(discontinuity, 3),
                   round(coloration, 3), round(loudness, 3)]
    }
    return table

with gr.Blocks(title="NISQA Speech Quality (MOS) Demo") as demo:
    gr.Markdown(
        """
        # 🎧 NISQA Speech Quality (MOS)  
        Upload or record speech and get **MOS + quality dimensions**.  
        Uses NISQA v2.0 via TorchMetrics.
        """
    )
    with gr.Row():
        audio = gr.Audio(sources=["mic", "upload"], type="filepath", label="Input audio (wav/mp3/m4a...)")
    btn = gr.Button("Predict")
    out = gr.Dataframe(headers=["Metric", "Score"], label="Results", interactive=False)

    btn.click(fn=predict_nisqa, inputs=audio, outputs=out)

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