Sanskrit_TTS_v2

Sanskrit Text-to-Speech Model

Model Overview

Model ID: R910/Sanskrit_TTS_v2
Base Model: unsloth/orpheus-3b-0.1-ft
Language: English Primary Dataset: IIT Madras IndicTTS Sanskrit Database
Voice: Mono Female
Training Audio Duration: 10.93 hours
This fine-tuned Language Model (LLaMA) specializes in Sanskrit text-to-speech synthesis and has been optimized using Unsloth and Hugging Face's TRL library for enhanced training efficiency.

Training Data

The model was trained on the Sanskrit speech corpus from the IIT Madras IndicTTS Database, using a mono female voice recording with a total audio duration of 10.93 hours. The IndicTTS project, developed by the Speech and Language Technology Group at IIT Madras, provides high-quality speech corpora for Indic languages.

Installation Requirements

Environment Detection and Base Setup

# Environment detection
python3 -c "
import os
print('colab' if 'COLAB_' in ''.join(os.environ.keys()) else 'local')
"

# Install core dependencies
pip install snac

Google Colab Installation

For Google Colab environments, execute the following installation sequence:

# Install Colab-specific dependencies
pip install --no-deps bitsandbytes accelerate xformers==0.0.29.post3 peft trl triton cut_cross_entropy unsloth_zoo
pip install sentencepiece protobuf 'datasets>=3.4.1,<4.0.0' huggingface_hub hf_transfer
pip install --no-deps unsloth
# Environment cleanup (recommended for clean installation)
pip uninstall torch torchvision torchaudio unsloth unsloth_zoo transformers -y
pip cache purge

# Install PyTorch with CUDA 12.1 support
pip install torch==2.4.1+cu121 torchvision==0.19.1+cu121 torchaudio==2.4.1+cu121 --index-url https://download.pytorch.org/whl/cu121

# Install latest Unsloth from source
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"

# Additional dependencies
pip install librosa
pip install -U datasets

Implementation Guide

Complete Implementation Code

import gradio as gr
import torch
from unsloth import FastLanguageModel
from IPython.display import display, Audio
import numpy as np

# Global model variables
model = None
tokenizer = None
snac_model = None
device = None
def load_models():
    """Initialize and load all required models for Sanskrit TTS inference."""
    global model, tokenizer, snac_model, device
    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Loading models on: {device}")
    
    # Load the fine-tuned Sanskrit TTS model
    model, tokenizer = FastLanguageModel.from_pretrained(
        "R910/Sanskrit_TTS_v2",  
        max_seq_length=2048,
        dtype=None,
        load_in_4bit=False,
    )
    
    model = model.to(device)
    FastLanguageModel.for_inference(model)
    
    # Load SNAC model for audio generation
    try:
        from snac import SNAC
        snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval()
    except ImportError:
        print("Warning: SNAC model import failed. Make sure SNAC is installed.")
    snac_model.to("cpu")
    
    print("Models loaded successfully!")
def redistribute_codes(code_list):
    """Redistribute generated codes into hierarchical layers for audio synthesis."""
    layer_1 = []
    layer_2 = []
    layer_3 = []
    
    for i in range((len(code_list)+1)//7):
        layer_1.append(code_list[7*i])
        layer_2.append(code_list[7*i+1]-4096)
        layer_3.append(code_list[7*i+2]-(2*4096))
        layer_3.append(code_list[7*i+3]-(3*4096))
        layer_2.append(code_list[7*i+4]-(4*4096))
        layer_3.append(code_list[7*i+5]-(5*4096))
        layer_3.append(code_list[7*i+6]-(6*4096))
    
    codes = [torch.tensor(layer_1).unsqueeze(0),
             torch.tensor(layer_2).unsqueeze(0),
             torch.tensor(layer_3).unsqueeze(0)]
    
    audio_hat = snac_model.decode(codes)
    return audio_hat
def sanskrit_tts_inference(sanskrit_text, chosen_voice=""):
    """
    Generate Sanskrit speech from input text using the fine-tuned model.
    
    Args:
        sanskrit_text (str): Input Sanskrit text in Devanagari script
        chosen_voice (str): Voice selection parameter (optional)
    
    Returns:
        tuple: (audio_data, status_message)
    """
    if not sanskrit_text.strip():
        return None, "Please enter some Sanskrit text."
    
    try:
        prompts = [sanskrit_text]
        chosen_voice = 1070
        
        # Prepare prompts with voice selection
        prompts_ = [(f"{chosen_voice}: " + p) if chosen_voice else p for p in prompts]
        
        # Tokenize input prompts
        all_input_ids = []
        for prompt in prompts_:
            input_ids = tokenizer(prompt, return_tensors="pt").input_ids
            all_input_ids.append(input_ids)
        
        # Define special tokens
        start_token = torch.tensor([[ 128259]], dtype=torch.int64)
        end_tokens = torch.tensor([[128009, 128260]], dtype=torch.int64) 
        
        # Construct modified input sequences
        all_modified_input_ids = []
        for input_ids in all_input_ids:
            modified_input_ids = torch.cat([start_token, input_ids, end_tokens], dim=1) 
            all_modified_input_ids.append(modified_input_ids)
        
        # Apply padding and create attention masks
        all_padded_tensors = []
        all_attention_masks = []
        max_length = max([modified_input_ids.shape[1] for modified_input_ids in all_modified_input_ids])
        
        for modified_input_ids in all_modified_input_ids:
            padding = max_length - modified_input_ids.shape[1]
            padded_tensor = torch.cat([torch.full((1, padding), 128263, dtype=torch.int64), modified_input_ids], dim=1)
            attention_mask = torch.cat([torch.zeros((1, padding), dtype=torch.int64), torch.ones((1, modified_input_ids.shape[1]), dtype=torch.int64)], dim=1)
            all_padded_tensors.append(padded_tensor)
            all_attention_masks.append(attention_mask)
        
        # Batch tensors for inference
        all_padded_tensors = torch.cat(all_padded_tensors, dim=0)
        all_attention_masks = torch.cat(all_attention_masks, dim=0)
        
        input_ids = all_padded_tensors.to(device)
        attention_mask = all_attention_masks.to(device)
        
        # Generate audio codes using the model
        generated_ids = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=1200,
            do_sample=True,
            temperature=0.6,
            top_p=0.95,
            repetition_penalty=1.1,
            num_return_sequences=1,
            eos_token_id=128258,
            use_cache=True
        )
        
        # Post-process generated tokens
        token_to_find = 128257
        token_to_remove = 128258
        
        token_indices = (generated_ids == token_to_find).nonzero(as_tuple=True)
        
        if len(token_indices[1]) > 0:
            last_occurrence_idx = token_indices[1][-1].item()
            cropped_tensor = generated_ids[:, last_occurrence_idx+1:]
        else:
            cropped_tensor = generated_ids
        
        mask = cropped_tensor != token_to_remove
        
        processed_rows = []
        for row in cropped_tensor:
            masked_row = row[row != token_to_remove]
            processed_rows.append(masked_row)
        
        # Convert tokens to audio codes
        code_lists = []
        for row in processed_rows:
            row_length = row.size(0)
            new_length = (row_length // 7) * 7
            trimmed_row = row[:new_length]
            trimmed_row = [t - 128266 for t in trimmed_row]
            code_lists.append(trimmed_row)
        
        # Generate audio samples
        my_samples = []
        for code_list in code_lists:
            samples = redistribute_codes(code_list)
            my_samples.append(samples)
        
        if len(my_samples) > 0:
            audio_sample = my_samples[0].detach().squeeze().to("cpu").numpy()
            return (24000, audio_sample), f"✅ Generated audio for: {sanskrit_text}"
        else:
            return None, "❌ Failed to generate audio - no valid codes produced."
            
    except Exception as e:
        return None, f"❌ Error during inference: {str(e)}"
# Initialize models
print("Loading models... This may take a moment.")
load_models()
# Create Gradio interface
with gr.Blocks(title="Sanskrit Text-to-Speech") as demo:
    gr.Markdown("""
    # 🕉️ Sanskrit Text-to-Speech
    
    Enter Sanskrit text in Devanagari script and generate speech using your fine-tuned model.
    """)
    
    with gr.Row():
        with gr.Column():
            sanskrit_input = gr.Textbox(
                label="Sanskrit Text",
                placeholder="Enter Sanskrit text in Devanagari script...",
                lines=3,
                value="नमस्ते" 
            )
            
            generate_btn = gr.Button("🎵 Generate Speech", variant="primary")
        
        with gr.Column():
            audio_output = gr.Audio(
                label="Generated Sanskrit Speech",
                type="numpy"
            )
            
            status_output = gr.Textbox(
                label="Status",
                lines=2,
                interactive=False
            )
    
    # Example inputs for demonstration
    gr.Examples(
        examples=[
            ["नमस्ते"],
            ["संस्कृत एक प्राचीन भाषा है"],
            ["ॐ शान्ति शान्ति शान्तिः"],
            ["सर्वे भवन्तु सुखिनः"],
        ],
        inputs=[sanskrit_input],
        outputs=[audio_output, status_output],
        fn=sanskrit_tts_inference,
        cache_examples=False
    )
    
    # Connect interface components
    generate_btn.click(
        fn=sanskrit_tts_inference,
        inputs=[sanskrit_input],
        outputs=[audio_output, status_output]
    )
# Launch the application
if __name__ == "__main__":
    demo.launch(
        share=True,  
        server_name="0.0.0.0",  
        server_port=7860,
        show_error=True
    )

🔊 Demo Outputs

<table> <tr> <td><strong>� यदा यदा हि धर्मस्य ग्लानिर्भवति भारत।</strong><br/><em>Bhagavad Gita 4.7</em></td> <td> <audio controls> <source src="https://huggingface.co/R910/Sanskrit_TTS_v2/resolve/main/यदा यदा हि धर्मस्य ग्लानिर्भवति भारत।.wav" type="audio/wav"> Your browser does not support the audio element. </audio> </td> </tr> <tr> <td><strong>🕉️ कर्मण्येवाधिकारस्ते मा फलेषु कदाचन।</strong><br/><em>Bhagavad Gita 2.47</em></td> <td> <audio controls> <source src="https://huggingface.co/R910/Sanskrit_TTS_v2/resolve/main/कर्मण्येवाधिकारस्ते मा फलेषु कदाचन।.wav" type="audio/wav"> Your browser does not support the audio element. </audio> </td> </tr> <tr> <td><strong>📚 विद्या ददाति विनयं</strong><br/><em>Subhashita</em></td> <td> <audio controls> <source src="https://huggingface.co/R910/Sanskrit_TTS_v2/resolve/main/विद्या ददाति विनयं.wav" type="audio/wav"> Your browser does not support the audio element. </audio> </td> </tr> <tr> <td><strong>🌟 तमसो मा ज्योतिर्गमय।</strong><br/><em>Brihadaranyaka Upanishad 1.3.28</em></td> <td> <audio controls> <source src="https://huggingface.co/R910/Sanskrit_TTS_v2/resolve/main/तमसो मा ज्योतिर्गमय।.wav" type="audio/wav"> Your browser does not support the audio element. </audio> </td> </tr> </table>

Model Information

Developer: R910
License: Apache 2.0
Base Architecture: Fine-tuned from unsloth/orpheus-3b-0.1-ft

This model has been optimized using Unsloth's efficient training framework, achieving 2x faster training speeds compared to standard implementations, in conjunction with Hugging Face's TRL (Transformer Reinforcement Learning) library.

Citation

If you use this model or the training data, please cite:

@inproceedings{indictts,
  title     = {Building Open Sourced and Industry Grade Low-Resource {TTS} for {I}ndian Languages},
  author    = {ID Prakashraj and Abhayjeet Singh and Anusha Prakash and AV Anand Kumar and Shambavi Bhaskar
               and Varun Srinivas and Vishal Sunder and Hema A Murthy and S Umesh},
  booktitle = {Proc. Interspeech 2023},
  year      = {2023},
  pages     = {1009--1013},
  doi       = {10.21437/Interspeech.2023-1339}
}

Dataset source: IIT Madras IndicTTS Database

Acknowledgments

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

Special thanks to the Speech and Language Technology Group, IIT Madras for providing the Sanskrit TTS dataset.

Technical Specifications

• Model Type: Fine-tuned Language Model for Text-to-Speech
• Architecture: LLaMA-based with LoRA adaptation
• Audio Output: 24kHz sampling rate
• Maximum Sequence Length: 2048 tokens
• Supported Script: Devanagari (Sanskrit)
• Training Framework: Unsloth + Hugging Face TRL

Usage Requirements

• Hardware: CUDA-compatible GPU
• Dependencies: PyTorch 2.4.1+, Transformers, SNAC audio codec
• Python Version: 3.7+