app.py

import gradio as gr
import torch
import torchaudio
from encoder.utils import convert_audio
from decoder.pretrained import WavTokenizer

# Initialize WavTokenizer
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
config_path = "wavtokenizer_smalldata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml"
model_path = "WavTokenizer_small_600_24k_4096.ckpt"

wavtokenizer = WavTokenizer.from_pretrained0802(config_path, model_path)
wavtokenizer = wavtokenizer.to(device)

def encode_audio(audio_file):
    # Load and preprocess the audio
    wav, sr = torchaudio.load(audio_file)
    wav = convert_audio(wav, sr, 24000, 1)
    wav = wav.to(device)
    
    # Encode the audio
    bandwidth_id = torch.tensor([0]).to(device)
    _, discrete_code = wavtokenizer.encode_infer(wav, bandwidth_id=bandwidth_id)
    
    # Convert the discrete code to a string representation
    code_str = ' '.join(map(str, discrete_code.cpu().numpy().flatten()))
    
    return code_str

# Create the Gradio interface
iface = gr.Interface(
    fn=encode_audio,
    inputs=gr.Audio(type="filepath"),
    outputs=gr.Textbox(label="Discrete Codes"),
    title="WavTokenizer Encoder Demo",
    description="Upload an audio file to see its WavTokenizer discrete codes. The output shows 40 tokens per second of audio."
)

# Launch the demo
iface.launch()