app.py

import os
import json
import gradio as gr
import shutil
import subprocess
import requests
import tarfile
from pathlib import Path
import soundfile as sf
import sherpa_onnx
from deep_translator import GoogleTranslator
import numpy as np
from iso639 import Lang
import pycountry

# Load model JSON
MODEL_JSON_URL = "https://github.com/willwade/tts-wrapper/blob/main/tts_wrapper/engines/sherpaonnx/merged_models.json"
MODEL_JSON_PATH = "./models.json"

# Load models
if not os.path.exists(MODEL_JSON_PATH):
    response = requests.get(MODEL_JSON_URL.replace("/blob/", "/raw/"))
    with open(MODEL_JSON_PATH, "w") as f:
        f.write(response.text)

with open(MODEL_JSON_PATH, "r") as f:
    models = json.load(f)

def get_model_display_info(model_info):
    """Create a display string for a model."""
    # Get language info
    lang_info = model_info.get('language', [{}])[0]
    lang_name = lang_info.get('language_name', lang_info.get('Language Name', 'Unknown'))
    lang_code = lang_info.get('lang_code', lang_info.get('Iso Code', 'Unknown'))
    
    # Get model info
    voice_name = model_info.get('name', model_info.get('id', 'Unknown'))
    developer = model_info.get('developer', '')
    quality = model_info.get('quality', 'MMS' if 'mms' in voice_name.lower() else '')
    
    # Create display name
    model_display = f"{voice_name} ({developer}"
    if quality:
        model_display += f" - {quality}"
    model_display += ")"
    
    # Combine language and model info
    return f"{lang_name} ({lang_code}) | {model_display}"

# Group models by language
models_by_lang = {}
for model_id, model_info in models.items():
    # Get language info from the first language in the list
    lang_info = model_info.get('language', [{}])[0]
    lang_name = lang_info.get('language_name', lang_info.get('Language Name', 'Unknown'))
    lang_code = lang_info.get('lang_code', lang_info.get('Iso Code', 'Unknown'))
    group_key = f"{lang_name} ({lang_code})"
    
    if group_key not in models_by_lang:
        models_by_lang[group_key] = []
    
    # Add model to language group
    models_by_lang[group_key].append((get_model_display_info(model_info), model_id))

# Create dropdown choices with model IDs as values
dropdown_choices = []
models_by_display = {}  # Map display names to model IDs
for lang, model_list in sorted(models_by_lang.items()):
    # Add all models in this language group
    for display_name, model_id in sorted(model_list):
        dropdown_choices.append(display_name)
        models_by_display[display_name] = model_id

def get_language_code(model_info):
    """Get the language code."""
    if not model_info.get("language"):
        return None
    
    lang_info = model_info["language"][0]
    # Try both key formats for language code
    lang_code = lang_info.get("lang_code", lang_info.get("Iso Code", "")).lower()
    return lang_code

# Special cases for codes not in ISO standard
SPECIAL_CODES = {
    "cmn": "zh",    # Mandarin Chinese
    "yue": "zh",    # Cantonese
    "pi": "el",     # Pali (using Greek for this model)
    "guj": "gu",    # Gujarati
}

def get_translate_code(iso_code):
    """Convert ISO code to Google Translate code."""
    if not iso_code:
        return None
        
    # Remove any script or dialect specifiers
    base_code = iso_code.split('-')[0].lower()
    
    # Check special cases first
    if base_code in SPECIAL_CODES:
        return SPECIAL_CODES[base_code]
        
    try:
        # Try to get the ISO 639-1 (2-letter) code
        lang = Lang(base_code)
        return lang.pt1
    except:
        # If that fails, try to find a matching language in pycountry
        try:
            lang = pycountry.languages.get(alpha_3=base_code)
            if lang and hasattr(lang, 'alpha_2'):
                return lang.alpha_2
        except:
            pass
    
    # If all else fails, try to use the original code
    if len(base_code) == 2:
        return base_code
    
    return None

def translate_text(input_text, source_lang="en", target_lang="en"):
    """Translate text using Google Translator."""
    try:
        # If source and target are the same, or if target is English, return as is
        if source_lang == target_lang or target_lang == "en":
            return input_text
            
        first_error = None
        # Try with original language code
        try:
            translated = GoogleTranslator(source=source_lang, target=target_lang).translate(input_text)
            return translated
        except Exception as e:
            first_error = e
            print(f"First translation attempt failed: {str(e)}")
            
        # Try with 'auto' as source language
        try:
            translated = GoogleTranslator(source='auto', target=target_lang).translate(input_text)
            return translated
        except Exception as e:
            print(f"Second translation attempt failed: {str(e)}")
            if first_error:
                raise first_error
            raise e
                
    except Exception as e:
        print(f"Translation error: {str(e)} for target language: {target_lang}")
        print(f"Attempted to use language code: {target_lang}")
        return input_text

def detect_language(text):
    """Detect the language of the input text using Google Translator."""
    try:
        return GoogleTranslator().detect(text).lower()
    except:
        return "en"  # Default to English if detection fails

def download_and_extract_model(url, destination):
    """Download and extract the model files."""
    print(f"Downloading from URL: {url}")
    print(f"Destination: {destination}")
    
    # Convert Hugging Face URL format if needed
    if "huggingface.co" in url:
        # Replace /tree/main/ with /resolve/main/ for direct file download
        base_url = url.replace("/tree/main/", "/resolve/main/")
        model_id = base_url.split("/")[-1]
        
        # Check if this is an MMS model
        is_mms_model = "mms-tts-multilingual-models-onnx" in url
        
        if is_mms_model:
            # MMS models have both model.onnx and tokens.txt
            model_url = f"{base_url}/model.onnx"
            tokens_url = f"{base_url}/tokens.txt"
            
            # Download model.onnx
            print("Downloading model.onnx...")
            model_path = os.path.join(destination, "model.onnx")
            response = requests.get(model_url, stream=True)
            if response.status_code != 200:
                raise Exception(f"Failed to download model from {model_url}. Status code: {response.status_code}")
            
            total_size = int(response.headers.get('content-length', 0))
            block_size = 8192
            downloaded = 0
            
            print(f"Total size: {total_size / (1024*1024):.1f} MB")
            with open(model_path, "wb") as f:
                for chunk in response.iter_content(chunk_size=block_size):
                    if chunk:
                        f.write(chunk)
                        downloaded += len(chunk)
                        if total_size > 0:
                            percent = int((downloaded / total_size) * 100)
                            if percent % 10 == 0:
                                print(f" {percent}%", end="", flush=True)
            print("\nModel download complete")
            
            # Download tokens.txt
            print("Downloading tokens.txt...")
            tokens_path = os.path.join(destination, "tokens.txt")
            response = requests.get(tokens_url, stream=True)
            if response.status_code != 200:
                raise Exception(f"Failed to download tokens from {tokens_url}. Status code: {response.status_code}")
            
            with open(tokens_path, "wb") as f:
                f.write(response.content)
            print("Tokens download complete")
            
            return
        else:
            # Other models are stored as tar.bz2 files
            url = f"{base_url}.tar.bz2"
        
        # Try the URL
        response = requests.get(url, stream=True)
        if response.status_code != 200:
            raise Exception(f"Failed to download model from {url}. Status code: {response.status_code}")
        
        # Check if this is a Git LFS file pointer
        content_start = response.content[:100].decode('utf-8', errors='ignore')
        if content_start.startswith('version https://git-lfs.github.com/spec/v1'):
            raise Exception(f"Received Git LFS pointer instead of file content from {url}")
    
    # Create model directory if it doesn't exist
    os.makedirs(destination, exist_ok=True)
    
    # For non-MMS models, handle tar.bz2 files
    tar_path = os.path.join(destination, "model.tar.bz2")
    
    # Download the file
    print("Downloading model archive...")
    response = requests.get(url, stream=True)
    total_size = int(response.headers.get('content-length', 0))
    block_size = 8192
    downloaded = 0
    
    print(f"Total size: {total_size / (1024*1024):.1f} MB")
    with open(tar_path, "wb") as f:
        for chunk in response.iter_content(chunk_size=block_size):
            if chunk:
                f.write(chunk)
                downloaded += len(chunk)
                if total_size > 0:
                    percent = int((downloaded / total_size) * 100)
                    if percent % 10 == 0:
                        print(f" {percent}%", end="", flush=True)
    print("\nDownload complete")
    
    # Extract the tar.bz2 file
    print(f"Extracting {tar_path} to {destination}")
    try:
        with tarfile.open(tar_path, "r:bz2") as tar:
            tar.extractall(path=destination)
        os.remove(tar_path)
        print("Extraction complete")
    except Exception as e:
        print(f"Error during extraction: {str(e)}")
        raise
    
    print("Contents of destination directory:")
    for root, dirs, files in os.walk(destination):
        print(f"\nDirectory: {root}")
        if dirs:
            print("  Subdirectories:", dirs)
        if files:
            print("  Files:", files)

def find_model_files(model_dir):
    """Find model files in the given directory and its subdirectories."""
    model_files = {}
    
    # Check if this is an MMS model
    is_mms = 'mms' in os.path.basename(model_dir).lower()
    
    for root, _, files in os.walk(model_dir):
        for file in files:
            file_path = os.path.join(root, file)
            
            # Model file
            if file.endswith('.onnx'):
                model_files['model'] = file_path
            
            # Tokens file
            elif file == 'tokens.txt':
                model_files['tokens'] = file_path
            
            # Lexicon file (only for non-MMS models)
            elif file == 'lexicon.txt' and not is_mms:
                model_files['lexicon'] = file_path
    
    # Create empty lexicon file if needed (only for non-MMS models)
    if not is_mms and 'model' in model_files and 'lexicon' not in model_files:
        model_dir = os.path.dirname(model_files['model'])
        lexicon_path = os.path.join(model_dir, 'lexicon.txt')
        with open(lexicon_path, 'w', encoding='utf-8') as f:
            pass  # Create empty file
        model_files['lexicon'] = lexicon_path
    
    return model_files if 'model' in model_files else {}

def generate_audio(text, model_info):
    """Generate audio from text using the specified model."""
    try:
        model_dir = os.path.join("./models", model_info['id'])
        
        print(f"\nLooking for model in: {model_dir}")
        
        # Download model if it doesn't exist
        if not os.path.exists(model_dir):
            print(f"Model directory doesn't exist, downloading {model_info['id']}...")
            os.makedirs(model_dir, exist_ok=True)
            download_and_extract_model(model_info['url'], model_dir)
        
        print(f"Contents of {model_dir}:")
        for item in os.listdir(model_dir):
            item_path = os.path.join(model_dir, item)
            if os.path.isdir(item_path):
                print(f"  Directory: {item}")
                print(f"    Contents: {os.listdir(item_path)}")
            else:
                print(f"  File: {item}")
        
        # Find and validate model files
        model_files = find_model_files(model_dir)
        if not model_files or 'model' not in model_files:
            raise ValueError(f"Could not find required model files in {model_dir}")
        
        print("\nFound model files:")
        print(f"Model: {model_files['model']}")
        print(f"Tokens: {model_files.get('tokens', 'Not found')}")
        print(f"Lexicon: {model_files.get('lexicon', 'Not required for MMS')}\n")
        
        # Check if this is an MMS model
        is_mms = 'mms' in os.path.basename(model_dir).lower()
        
        # Create configuration based on model type
        if is_mms:
            if 'tokens' not in model_files or not os.path.exists(model_files['tokens']):
                raise ValueError("tokens.txt is required for MMS models")
                
            # MMS models use tokens.txt and no lexicon
            vits_config = sherpa_onnx.OfflineTtsVitsModelConfig(
                model_files['model'],  # model
                '',                    # lexicon
                model_files['tokens'], # tokens
                '',                    # data_dir
                '',                    # dict_dir
                0.667,                 # noise_scale
                0.8,                   # noise_scale_w
                1.0                    # length_scale
            )
        else:
            # Non-MMS models use lexicon.txt
            if 'tokens' not in model_files or not os.path.exists(model_files['tokens']):
                raise ValueError("tokens.txt is required for VITS models")
                
            # Set data dir if it exists
            espeak_data = os.path.join(os.path.dirname(model_files['model']), 'espeak-ng-data')
            data_dir = espeak_data if os.path.exists(espeak_data) else ''
            
            # Get lexicon path if it exists
            lexicon = model_files.get('lexicon', '') if os.path.exists(model_files.get('lexicon', '')) else ''
            
            # Create VITS model config
            vits_config = sherpa_onnx.OfflineTtsVitsModelConfig(
                model_files['model'],  # model
                lexicon,               # lexicon
                model_files['tokens'], # tokens
                data_dir,             # data_dir
                '',                   # dict_dir
                0.667,                # noise_scale
                0.8,                  # noise_scale_w
                1.0                   # length_scale
            )
        
        # Create the model config with VITS
        model_config = sherpa_onnx.OfflineTtsModelConfig()
        model_config.vits = vits_config
        
        # Create TTS configuration
        config = sherpa_onnx.OfflineTtsConfig(
            model=model_config,
            max_num_sentences=2
        )
        
        # Initialize TTS engine
        tts = sherpa_onnx.OfflineTts(config)

        # Generate audio
        audio_data = tts.generate(text)

        # Ensure we have valid audio data
        if audio_data is None or len(audio_data.samples) == 0:
            raise ValueError("Failed to generate audio - no data generated")
            
        # Convert samples list to numpy array and normalize
        audio_array = np.array(audio_data.samples, dtype=np.float32)
        if np.any(audio_array):  # Check if array is not all zeros
            audio_array = audio_array / np.abs(audio_array).max()
        else:
            raise ValueError("Generated audio is empty")
            
        # Return in Gradio's expected format (numpy array, sample rate)
        return (audio_array, audio_data.sample_rate)
            
    except Exception as e:
        error_msg = str(e)
        # Check for OOV or token conversion errors
        if "out of vocabulary" in error_msg.lower() or "token" in error_msg.lower():
            error_msg = f"Text contains unsupported characters: {error_msg}"
        print(f"Error generating audio: {error_msg}")
        print(f"Error in TTS generation: {error_msg}")
        raise

def tts_interface(selected_model, text, translate_enabled, status_output):
    try:
        if not text.strip():
            return None, "Please enter some text"
            
        # Get model ID from the display name mapping
        model_id = models_by_display.get(selected_model)
        if not model_id or model_id not in models:
            return None, "Please select a model"
            
        model_info = models[model_id]
        
        # Get the language code and check if translation is needed
        lang_code = get_language_code(model_info)
        translate_code = get_translate_code(lang_code)
        
        # Store original text for status message
        original_text = text
        translated_text = None
        was_translated = False
        
        # Only translate if translation is enabled and needed
        if translate_enabled and translate_code and translate_code != "en":
            if not translate_code:
                return None, f"Cannot determine translation target language from code: {lang_code}"
                
            print(f"Translating to {translate_code}")
            translated_text = translate_text(text, "en", translate_code)
            if translated_text != text:  # Only mark as translated if text actually changed
                was_translated = True
                text = translated_text
        
        try:
            # Update status with language info
            lang_info = model_info.get('language', [{}])[0]
            lang_name = lang_info.get('language_name', 'Unknown')
            voice_name = model_info.get('name', model_id)
            status = f"Generating speech using {voice_name} ({lang_name})..."
            
            # Generate audio
            audio_data, sample_rate = generate_audio(text, model_info)
            
            # Include translation info in final status if text was actually translated
            final_status = f"Generated speech using {voice_name} ({lang_name})"
            if was_translated:
                final_status += f"\nTranslated: '{original_text}' → '{translated_text}'"
            else:
                final_status += f"\nText: '{text}'"
            
            return (sample_rate, audio_data), final_status
        except ValueError as e:
            # Handle known errors with user-friendly messages
            error_msg = str(e)
            if "cannot process some words" in error_msg.lower():
                return None, error_msg
            return None, f"Error: {error_msg}"
            
    except Exception as e:
        print(f"Error in TTS generation: {str(e)}")
        error_msg = str(e)
        return None, f"Error: {error_msg}"

# Gradio Interface
with gr.Blocks() as app:
    gr.Markdown("# Sherpa-ONNX TTS Demo")
    with gr.Row():
        with gr.Column():
            model_dropdown = gr.Dropdown(
                choices=dropdown_choices,
                label="Select Model",
                value=dropdown_choices[0] if dropdown_choices else None
            )
            text_input = gr.Textbox(
                label="Text to speak",
                placeholder="Enter text here...",
                lines=3
            )
            translate_checkbox = gr.Checkbox(
                label="Translate to model language",
                value=False
            )
            with gr.Row():
                generate_btn = gr.Button("Generate Audio")
                stop_btn = gr.Button("Stop")
            
        with gr.Column():
            audio_output = gr.Audio(
                label="Generated Audio",
                type="numpy"
            )
            status_text = gr.Textbox(
                label="Status",
                interactive=False
            )
    
    def on_model_change(model_name):
        # Get model info
        model_id = models_by_display.get(model_name)
        if not model_id or model_id not in models:
            return False
            
        model_info = models[model_id]
        lang_code = get_language_code(model_info)
        
        # Auto-check translation for non-English models
        should_translate = lang_code and lang_code.lower() != "en"
        return should_translate

    # Update translation checkbox when model changes
    model_dropdown.change(
        fn=on_model_change,
        inputs=[model_dropdown],
        outputs=[translate_checkbox]
    )

    # Set up event handlers
    gen_event = generate_btn.click(
        fn=tts_interface,
        inputs=[model_dropdown, text_input, translate_checkbox, status_text],
        outputs=[audio_output, status_text]
    )
    
    stop_btn.click(
        fn=None,
        cancels=gen_event,
        queue=False
    )

app.launch()