app.py

import os
import re
import json
import torch
import inflect
import random
import uroman as ur
import numpy as np
import torchaudio
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from outetts.wav_tokenizer.decoder import WavTokenizer
from yarngpt.audiotokenizer import AudioTokenizerV2

# Initialize paths and models
tokenizer_path = "saheedniyi/YarnGPT2"
wav_tokenizer_config_path = "wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml"
wav_tokenizer_model_path = "wavtokenizer_large_speech_320_24k.ckpt"

# Initialize the audio tokenizer
audio_tokenizer = AudioTokenizerV2(
    tokenizer_path, wav_tokenizer_model_path, wav_tokenizer_config_path
)

# Load the model
model = AutoModelForCausalLM.from_pretrained(
    tokenizer_path, torch_dtype="auto"
).to(audio_tokenizer.device)

# Function to generate speech
def generate_speech(text, language, speaker_name, temperature=0.1, repetition_penalty=1.1):
    # Create prompt
    prompt = audio_tokenizer.create_prompt(text, lang=language, speaker_name=speaker_name)
    
    # Tokenize prompt
    input_ids = audio_tokenizer.tokenize_prompt(prompt)
    
    # Generate output
    output = model.generate(
        input_ids=input_ids,
        temperature=temperature,
        repetition_penalty=repetition_penalty,
        max_length=4000,
    )
    
    # Get audio codes and convert to audio
    codes = audio_tokenizer.get_codes(output)
    audio = audio_tokenizer.get_audio(codes)
    
    # Save audio to file
    output_path = "output.wav"
    torchaudio.save(output_path, audio, sample_rate=24000)
    
    return output_path

# Create Gradio interface
def tts_interface(text, language, speaker_name, temperature, repetition_penalty):
    try:
        audio_path = generate_speech(
            text, 
            language, 
            speaker_name,
            temperature,
            repetition_penalty
        )
        return audio_path
    except Exception as e:
        return f"Error: {str(e)}"

# Define available languages and speakers
languages = ["english", "igbo", "yoruba", "hausa", "pidgin"]
speakers = ["idera", "enitan", "abeo", "eniola", "kachi", "aisha", "amara", "bello", "chidi"]

# Create the Gradio interface
demo = gr.Interface(
    fn=tts_interface,
    inputs=[
        gr.Textbox(label="Text to convert to speech", lines=5),
        gr.Dropdown(languages, label="Language", value="english"),
        gr.Dropdown(speakers, label="Speaker", value="idera"),
        gr.Slider(0.1, 1.0, value=0.1, label="Temperature"),
        gr.Slider(1.0, 2.0, value=1.1, label="Repetition Penalty"),
    ],
    outputs=gr.Audio(type="filepath"),
    title="YarnGPT Text-to-Speech",
    description="Convert text to speech using YarnGPT model for various African languages",
)

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