Jeyaram-K/piper-tamil-voices

This repository provides high-quality, neural Tamil Text-to-Speech (TTS) models trained specifically for the Piper TTS engine. Built on the robust VITS architecture and exported to ONNX format, these fast and efficient voice synthesis models are optimized for local inference, accessibility tools, smart assistants, and any application requiring natural-sounding Tamil language generation.

Models Available

Voice	Language	Quality	Format
ta_IN-ValluvaNeural-medium	Tamil (India)	Medium	ONNX

Usage

You can use these voices with Piper TTS. Replace <model_name> with the name of the voice model you wish to use (e.g., ta_IN-ValluvaNeural-medium).

Command Line

echo "வணக்கம், எப்படி இருக்கிறீர்கள்?" | piper \
  --model <model_name>/<model_name>.onnx.onnx \
  --config <model_name>/<model_name>.onnx.json \
  --output_file output.wav

Example using ta_IN-ValluvaNeural-medium:

echo "வணக்கம், எப்படி இருக்கிறீர்கள்?" | piper \
  --model ta_IN-ValluvaNeural-medium/ta_IN-ValluvaNeural-medium.onnx.onnx \
  --config ta_IN-ValluvaNeural-medium/ta_IN-ValluvaNeural-medium.onnx.json \
  --output_file output.wav

Python

import wave
from piper import PiperVoice

# Replace with your chosen model name
model_name = "ta_IN-ValluvaNeural-medium"
model_path = f"{model_name}/{model_name}.onnx.onnx"
config_path = f"{model_name}/{model_name}.onnx.json"

voice = PiperVoice.load(model_path, config_path=config_path)

text = "வணக்கம், எப்படி இருக்கிறீர்கள்?"

with wave.open("output.wav", "wb") as wav_file:
    voice.synthesize(text, wav_file)

Details

Each model directory contains the ONNX model files, the training checkpoint, and the dataset used for training that specific model. For example, ta_IN-ValluvaNeural-medium was trained up to 2204 epochs and 1388260 steps.

File Structure

The repository is structured with one directory per model:

• <model_name>/
  • <model_name>.onnx.onnx: The exported ONNX model.
  • <model_name>.onnx.json: The model configuration file.
  • *.ckpt: The training checkpoint file.
  • *.jsonl: The dataset used for training.

Model Details

Intended Use

These models are intended for text-to-speech applications where natural-sounding Tamil voices are required. Suitable for accessibility tools, educational applications, voice assistants, and audio generation.

Limitations

• The models may struggle with non-Tamil words, specialized jargon, or heavy mixed-language text (e.g., Tanglish).
• Prosody and emotion are limited to the specific speaking style present in the training datasets.

Ethical Considerations

Please use these synthetic voices responsibly. Ensure that the use of these voices complies with local laws and ethical standards regarding synthetic media. Do not use for impersonation, generating misleading content, or any harmful purposes.