app.py

# pip install gradio torch torchaudio snac

import torch
import torchaudio
import gradio as gr
from snac import SNAC

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
TARGET_SR = 32000  # using the 32 kHz model per your example
MODEL = SNAC.from_pretrained("hubertsiuzdak/snac_32khz").eval().to(DEVICE)

def encode_then_decode(audio_in):
    if audio_in is None:
        return None

    sr, data = audio_in  # data: (T,) mono or (T, C) stereo

    # mono-ize if needed
    if data.ndim == 2:
        data = data.mean(axis=1)

    # torchify to [1, T]
    x = torch.from_numpy(data).float().unsqueeze(0)

    # resample to model's target SR
    if sr != TARGET_SR:
        x = torchaudio.functional.resample(x, orig_freq=sr, new_freq=TARGET_SR)

    # expand to [B, 1, T] then encode->decode
    x = x.unsqueeze(0).to(DEVICE)  # [1, 1, T]
    with torch.inference_mode():
        codes = MODEL.encode(x)
        y = MODEL.decode(codes)  # [1, 1, T]

    y = y.squeeze().detach().cpu().numpy()
    return (TARGET_SR, y)

with gr.Blocks(title="SNAC Encode→Decode (Simple)") as demo:
    gr.Markdown("## 🎧 SNAC Encode → Decode (32 kHz)\nResample → `encode()` → `decode()` — that’s it.")
    with gr.Row():
        with gr.Column():
            audio_in = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Input audio")
            run = gr.Button("Encode + Decode")
        with gr.Column():
            audio_out = gr.Audio(type="numpy", label="Reconstructed (32 kHz)")
    run.click(encode_then_decode, inputs=audio_in, outputs=audio_out)

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