app.py

import gradio as gr
import torch
import torchaudio
import os
import re
import subprocess
from transformers import AutoModelForCausalLM
from yarngpt_utils import AudioTokenizer

# Download model files if they don't exist
def download_if_not_exists(url, filename):
    if not os.path.exists(filename):
        print(f"Downloading {filename}...")
        subprocess.run(["wget", url, "-O", filename])
        print(f"Downloaded {filename}")

# Download necessary files
download_if_not_exists(
    "https://huggingface.co/novateur/WavTokenizer-medium-speech-75token/resolve/main/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml",
    "wavtokenizer_config.yaml"
)
download_if_not_exists(
    "https://huggingface.co/novateur/WavTokenizer-large-speech-75token/blob/main/wavtokenizer_large_speech_320_v2.ckpt",
    "wavtokenizer_model.ckpt"
)

# Initialize the model (this runs when the app starts)
def initialize_model():
    # Set paths
    hf_path = "saheedniyi/YarnGPT"
    wav_tokenizer_config_path = "wavtokenizer_config.yaml"
    wav_tokenizer_model_path = "wavtokenizer_model.ckpt"
    
    # Create AudioTokenizer
    audio_tokenizer = AudioTokenizer(
        hf_path, wav_tokenizer_model_path, wav_tokenizer_config_path
    )
    
    # Load model
    model = AutoModelForCausalLM.from_pretrained(hf_path, torch_dtype="auto").to(audio_tokenizer.device)
    
    return model, audio_tokenizer

# Generate audio from text
def generate_speech(text, speaker_name):
    # Create prompt
    prompt = audio_tokenizer.create_prompt(text, speaker_name)
    
    # Tokenize prompt
    input_ids = audio_tokenizer.tokenize_prompt(prompt)
    
    # Generate output
    output = model.generate(
        input_ids=input_ids,
        temperature=0.1,
        repetition_penalty=1.1,
        max_length=4000,
    )
    
    # Convert to audio codes
    codes = audio_tokenizer.get_codes(output)
    
    # Convert codes to audio
    audio = audio_tokenizer.get_audio(codes)
    
    # Save audio temporarily
    temp_path = "output.wav"
    torchaudio.save(temp_path, audio, sample_rate=24000)
    
    return temp_path

# Load model globally
print("Loading model...")
model, audio_tokenizer = initialize_model()
print("Model loaded!")

# Create Gradio interface
speakers = ["idera", "emma", "jude", "osagie", "tayo", "zainab", "joke", "regina", "remi", "umar", "chinenye"]

demo = gr.Interface(
    fn=generate_speech,
    inputs=[
        gr.Textbox(lines=5, placeholder="Enter text here..."),
        gr.Dropdown(choices=speakers, label="Speaker", value="idera")
    ],
    outputs=gr.Audio(type="filepath"),
    title="YarnGPT: Nigerian Accented Text-to-Speech",
    description="Generate natural-sounding Nigerian accented speech from text."
)

demo.launch()