app.py

import os
import sys
import gradio as gr
import torch
import torchaudio
import uroman
import numpy as np
from transformers import AutoModelForCausalLM, AutoTokenizer
from outetts.wav_tokenizer.decoder import WavTokenizer

# Clone and install YarnGPT at startup
if not os.path.exists("yarngpt"):
    print("Cloning YarnGPT repository...")
    os.system("git clone https://github.com/saheedniyi02/yarngpt.git")
    # Add the repository to Python path
    sys.path.append("yarngpt")

# Import the YarnGPT AudioTokenizer
from yarngpt.audiotokenizer import AudioTokenizerV2

# Constants and paths
MODEL_PATH = "saheedniyi/YarnGPT2b"
WAV_TOKENIZER_CONFIG_PATH = "wavtokenizer_config.yaml"
WAV_TOKENIZER_MODEL_PATH = "wavtokenizer_model.ckpt"

# Download the model files at startup
if not os.path.exists(WAV_TOKENIZER_CONFIG_PATH):
    print("Downloading WavTokenizer config...")
    os.system(f"wget -O {WAV_TOKENIZER_CONFIG_PATH} https://huggingface.co/novateur/WavTokenizer-medium-speech-75token/resolve/main/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml")

if not os.path.exists(WAV_TOKENIZER_MODEL_PATH):
    print("Downloading WavTokenizer model...")
    os.system(f"wget -O {WAV_TOKENIZER_MODEL_PATH} https://huggingface.co/novateur/WavTokenizer-large-speech-75token/resolve/main/wavtokenizer_large_speech_320_24k.ckpt")

# Initialize the model and tokenizer
def initialize_model():
    print("Initializing AudioTokenizer and model...")
    audio_tokenizer = AudioTokenizerV2(
        MODEL_PATH, 
        WAV_TOKENIZER_MODEL_PATH, 
        WAV_TOKENIZER_CONFIG_PATH
    )
    
    print("Loading YarnGPT model...")
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_PATH, 
        torch_dtype="auto"
    ).to(audio_tokenizer.device)
    
    return model, audio_tokenizer

# Initialize the model and tokenizer
print("Starting model initialization...")
model, audio_tokenizer = initialize_model()
print("Model initialization complete!")

# Available voices and languages
VOICES = ["idera", "jude", "kemi", "tunde", "funmi"]
LANGUAGES = ["english", "yoruba", "igbo", "hausa", "pidgin"]

# Function to generate speech
def generate_speech(text, language, voice, temperature=0.1, rep_penalty=1.1):
    if not text:
        return None, "Please enter some text to convert to speech."
    
    try:
        # Create prompt
        prompt = audio_tokenizer.create_prompt(text, lang=language, speaker_name=voice)
        
        # Tokenize prompt
        input_ids = audio_tokenizer.tokenize_prompt(prompt)
        
        # Generate output
        output = model.generate(
            input_ids=input_ids,
            temperature=temperature,
            repetition_penalty=rep_penalty,
            max_length=4000,
        )
        
        # Convert to audio
        codes = audio_tokenizer.get_codes(output)
        audio = audio_tokenizer.get_audio(codes)
        
        # Save audio to file
        temp_audio_path = "output.wav"
        torchaudio.save(temp_audio_path, audio, sample_rate=24000)
        
        return temp_audio_path, f"Successfully generated speech for: {text[:50]}..."
    
    except Exception as e:
        return None, f"Error generating speech: {str(e)}"

# Example text for demonstration
examples = [
    ["Hello, my name is Claude. I am an AI assistant created by Anthropic.", "english", "idera"],
    ["Báwo ni o ṣe wà? Mo ń gbádùn ọjọ́ mi.", "yoruba", "kemi"],
    ["I don dey come house now, make you prepare food.", "pidgin", "jude"]
]

# Create the Gradio interface
with gr.Blocks(title="YarnGPT - Nigerian Accented Text-to-Speech") as demo:
    gr.Markdown("# YarnGPT - Nigerian Accented Text-to-Speech")
    gr.Markdown("Generate speech with Nigerian accents using YarnGPT model.")
    
    with gr.Tab("Basic TTS"):
        with gr.Row():
            with gr.Column():
                text_input = gr.Textbox(
                    label="Text to convert to speech", 
                    placeholder="Enter text here...",
                    lines=5
                )
                language = gr.Dropdown(
                    label="Language", 
                    choices=LANGUAGES, 
                    value="english"
                )
                voice = gr.Dropdown(
                    label="Voice", 
                    choices=VOICES, 
                    value="idera"
                )
                temperature = gr.Slider(
                    label="Temperature", 
                    minimum=0.1, 
                    maximum=1.0, 
                    value=0.1, 
                    step=0.1
                )
                rep_penalty = gr.Slider(
                    label="Repetition Penalty", 
                    minimum=1.0, 
                    maximum=2.0, 
                    value=1.1, 
                    step=0.1
                )
                generate_btn = gr.Button("Generate Speech")
            
            with gr.Column():
                audio_output = gr.Audio(label="Generated Speech")
                status_output = gr.Textbox(label="Status")
    
        gr.Examples(
            examples=examples,
            inputs=[text_input, language, voice],
            outputs=[audio_output, status_output],
            fn=generate_speech,
            cache_examples=False
        )
    
    generate_btn.click(
        generate_speech, 
        inputs=[text_input, language, voice, temperature, rep_penalty], 
        outputs=[audio_output, status_output]
    )
    
    gr.Markdown("""
    ## About YarnGPT
    YarnGPT is a text-to-speech model with Nigerian accents. It supports multiple languages and voices.
    
    ### Credits
    - Model by [saheedniyi](https://huggingface.co/saheedniyi/YarnGPT2b)
    - [Original Repository](https://github.com/saheedniyi02/yarngpt)
    """)

# Launch the app
if __name__ == "__main__":
    demo.launch()