Modelpublic

chan4lk/piper-tts-sinhala

sourceHugging Facemitupdated 6mo agoView on Hugging Face

0likes137downloads

Model Card

Piper TTS — Sinhala (si_LK) Medium

A single-speaker Sinhala text-to-speech model trained with Piper using the VITS architecture.

Model Details

Property	Value
Language	Sinhala (si)
Region	Sri Lanka (LK)
Quality	Medium
Architecture	VITS
Sample Rate	22050 Hz
Speakers	1
Phonemizer	eSpeak-ng
Format	ONNX
Model Size	~60 MB

Training Data

Trained on OpenSLR 30 — Google's high-quality Sinhala multi-speaker TTS corpus.

Quick Start

CLI

bash

pip install piper-tts

echo 'ආයුබෝවන්, මම සිංහලෙන් කතා කරමි.' | piper \
  --model si_LK-sinhala-medium.onnx \
  --output_file output.wav

Python

python

from piper import PiperVoice
import wave

voice = PiperVoice.load(
    "si_LK-sinhala-medium.onnx",
    config_path="si_LK-sinhala-medium.onnx.json"
)

with wave.open("output.wav", "w") as wav_file:
    voice.synthesize("ආයුබෝවන්, මම සිංහලෙන් කතා කරමි.", wav_file)

Batch Script

A convenience script is included for converting text files:

bash

python sinhala_tts.py input.txt -o output.wav
python sinhala_tts.py input.txt --speed 0.9 --silence 0.5
echo "ආයුබෝවන්" | python sinhala_tts.py -

Sample Outputs

Text	Audio
ආයුබෝවන්, මම සිංහලෙන් කතා කරමි.	sample.wav

Files

File	Description
`si_LK-sinhala-medium.onnx`	ONNX model weights
`si_LK-sinhala-medium.onnx.json`	Model config (phoneme map, inference params)
`sinhala_tts.py`	Batch TTS script for text files
`MODEL_CARD.md`	Detailed model card

Synthesis Parameters

Parameter	Default	Description
`noise_scale`	0.667	Controls variation/expressiveness
`length_scale`	1.0	Controls speed (lower = faster)
`noise_w`	0.8	Controls phoneme duration variation

Requirements

—piper-tts (pip install)
—espeak-ng (system package — provides Sinhala G2P)

Training

Trained on Google Colab (T4 GPU). Training notebook: piper-tts-sinhala-training.ipynb

License

MIT