abhishekdaygavhane/MarathiNaad-TTS

MarathiNaad‑TTS 🎙️

Overview

MarathiNaad‑TTS is a high‑quality text‑to‑speech model for the Marathi language. It uses the KittenTTS library (a lightweight TTS engine) with an ONNX‑exported model for fast inference.

Demo

Installation

git clone https://huggingface.co/your-username/MarathiNaad-TTS
cd MarathiNaad-TTS
pip install -r requirements.txt

Quick inference

from marathitts import KittenTTS

# Load the model (local repo directory)
model = KittenTTS(local_dir=".")

# Generate speech
audio = model.generate("नमस्कार, आज आपण एका नवीन विषयवार चर्चा करणार आहोत .", voice="mf_asha", speed=1.0)

# Save to file
import soundfile as sf
sf.write("output.wav", audio, 24000)

Repository contents

File	Size	Description
model.onnx	120 MB	Inference‑ready ONNX model
config.json	4 KB	Model configuration (sampling rate, phoneme set)
vocab.txt	2 KB	Token‑to‑phoneme mapping
README.md	—	This file
requirements.txt	—	Python dependencies

Training data & method

• Dataset: ~10 h of studio‑recorded Marathi speech (public domain + licensed recordings).
• Pre‑processing: Normalized amplitude, removed silences, extracted phonemes with phonemizer.
• Model: Fine‑tuned Tacotron‑2 + WaveRNN pipeline, exported to ONNX for fast CPU inference.

Evaluation

• Mean Opinion Score (MOS): 4.2 ± 0.3 (based on 30 native speakers).
• Sample scripts: evaluate.py (generates MOS on a held‑out set).

Limitations & Bias

• Trained only on neutral‑tone speech; may not handle extreme emotions well.
• Accent coverage limited to central Marathi dialects.
• Not intended for hateful, abusive, or disallowed content.

License

This model is released under the Apache‑2.0 License. See LICENSE for details.

Citation

If you use this model in research, please cite:

@misc{MarathiNaadTTS,
  author = Abhishek Daygavhane,
  title = MarathiNaad‑TTS: Marathi Text‑to‑Speech using KittenTTS,
  year = 2026,
  url = https://huggingface.co/your-username/MarathiNaad-TTS
}

Acknowledgements

• KittenTTS library for the inference engine.
• Coqui TTS for the training pipeline.
• ONNX Runtime for fast inference.
• Dataset contributors (list or link).

References

• MarathiNaad‑TTS repository: https://huggingface.co/your-username/MarathiNaad-TTS
• KittenTTS library: https://github.com/your-username/KittenTTS
• Coqui TTS paper: https://arxiv.org/abs/2106.05822
• ONNX Runtime documentation: https://onnxruntime.ai
• Mozilla Common Voice Marathi dataset: https://commonvoice.mozilla.org/mr

Questions / Issues

Open an issue on this repository or contact abhishekdaygavhane@gmail.com.