app.py

import subprocess
import sys
import os
import tempfile
from huggingface_hub import hf_hub_download

# --- 1. PRE-FLIGHT: BYPASS BUILD ISOLATION ---
def pre_flight_setup():
    try:
        import chatterbox
    except ImportError:
        print("Applying pkuseg build isolation bypass...")
        subprocess.check_call([sys.executable, "-m", "pip", "install", "numpy<2.0.0", "cython", "wheel"])
        subprocess.check_call([sys.executable, "-m", "pip", "install", "--no-build-isolation", "pkuseg==0.0.25"])
        subprocess.check_call([sys.executable, "-m", "pip", "install", "chatterbox-tts>=0.1.7"])

pre_flight_setup()

# --- 2. MAIN APPLICATION ---
import gradio as gr
import torch
import torch.nn as nn
import torchaudio as ta
from peft import PeftModel

from chatterbox.tts import ChatterboxTTS
from chatterbox.models.tokenizers import EnTokenizer

device = "cuda" if torch.cuda.is_available() else "cpu"
repo_id = "Praha-Labs/PrahaTTS-ML"

def load_model():
    print(f"Loading base Chatterbox model on {device}...")
    model = ChatterboxTTS.from_pretrained(device=device)

    print("Applying custom Indic tokenizer...")
    try:
        tokenizer_path = hf_hub_download(repo_id=repo_id, filename="tokenizer_indic.json")
        model.tokenizer = EnTokenizer(tokenizer_path)
    except Exception as e:
        print(f"Error during tokenizer inject: {e}")

    # --- CRITICAL FIX: MANUALLY RESIZE PYTORCH EMBEDDINGS ---
    # We must resize the base model's vocabulary layers to match the new 
    # Malayalam vocab size (2573) before loading the adapter weights.
    vocab_size = 2573
    print(f"Resizing base model embeddings to handle vocab size of {vocab_size}...")
    
    target_layer = model.t3 if hasattr(model, 't3') else model
    
    if hasattr(target_layer, 'text_emb'):
        embed_dim = target_layer.text_emb.embedding_dim
        target_layer.text_emb = nn.Embedding(vocab_size, embed_dim)
        
    if hasattr(target_layer, 'text_head'):
        in_features = target_layer.text_head.in_features
        has_bias = target_layer.text_head.bias is not None
        target_layer.text_head = nn.Linear(in_features, vocab_size, bias=has_bias)
        
    # Send resized layers to the correct device
    target_layer.to(device)

    print("Loading LoRA adapter weights...")
    try:
        if hasattr(model, 't3'):
            model.t3 = PeftModel.from_pretrained(model.t3, repo_id)
        else:
            model = PeftModel.from_pretrained(model, repo_id)
        print("LoRA adapter loaded successfully.")
    except Exception as e:
        print(f"Failed to load PEFT adapter: {e}")
        
    return model

# Initialize Model
tts_model = load_model()

def synthesize_audio(text, ref_audio, exaggeration, cfg_weight):
    if not text.strip():
        return None, "Please enter some text."
        
    audio_prompt_path = ref_audio if ref_audio else None
    
    try:
        wav = tts_model.generate(
            text, 
            audio_prompt_path=audio_prompt_path,
            exaggeration=exaggeration,
            cfg_weight=cfg_weight
        )
        
        temp_out = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
        ta.save(temp_out.name, wav.cpu(), tts_model.sr)
        return temp_out.name, "Generation successful!"
        
    except Exception as e:
        return None, f"Generation Error: {str(e)}"

# Define the Gradio Interface
with gr.Blocks(title="PrahaTTS-ML: Malayalam TTS", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🗣️ PrahaTTS-ML: Malayalam LoRA Adapter")
    gr.Markdown(
        "This Space runs the [Praha-Labs/PrahaTTS-ML](https://huggingface.co/Praha-Labs/PrahaTTS-ML) model. "
        "It is a Malayalam LoRA adapter built on top of ResembleAI's Chatterbox non-turbo TTS model. \n\n"
        "**Note**: Provide up to 5-10 seconds of clear reference audio for voice cloning capabilities."
    )
    
    with gr.Row():
        with gr.Column():
            text_input = gr.Textbox(
                label="Input Text (Malayalam)", 
                lines=4, 
                placeholder="നമസ്കാരം, നിങ്ങൾക്കെങ്ങനെയുണ്ട്?"
            )
            ref_audio_input = gr.Audio(
                label="Reference Voice Audio (Optional, for Voice Cloning)", 
                type="filepath"
            )
            
            with gr.Accordion("Advanced Voice Controls", open=False):
                exaggeration_slider = gr.Slider(
                    minimum=0.0, maximum=1.0, value=0.5, step=0.05, 
                    label="Emotion Exaggeration", 
                    info="Lower for monotone, higher for dramatic/expressive"
                )
                cfg_slider = gr.Slider(
                    minimum=0.0, maximum=1.0, value=0.5, step=0.05, 
                    label="CFG Weight", 
                    info="Lower if speech is too fast, higher to strictly mimic the reference voice"
                )
                
            generate_btn = gr.Button("Synthesize Speech", variant="primary")
            
        with gr.Column():
            audio_output = gr.Audio(label="Generated Output", interactive=False)
            status_output = gr.Textbox(label="Status Logging", interactive=False)
            
    generate_btn.click(
        fn=synthesize_audio,
        inputs=[text_input, ref_audio_input, exaggeration_slider, cfg_slider],
        outputs=[audio_output, status_output]
    )

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