An open-source text-to-speech project for the Chechen language (нохчийн мотт).
Chechen is a Nakh-Dagestanian language spoken by around 1.5 million people. Despite its size, it is not supported by major TTS providers such as Google, Amazon, ElevenLabs, Yandex, or Sber SaluteSpeech.
This repository provides a working baseline for Chechen text-to-speech based on Meta's MMS-TTS-che model. The current version is functional but far from perfect — pronunciation and intonation are limited.
A dedicated, fine-tuned model with significantly better quality is currently in development and will be released here as soon as it is ready.
- Python 3.9 or newer
- ~2 GB of free disk space (for dependencies and the model)
- Internet connection on first run (to download the model)
- Works on CPU; a GPU makes it faster but is not required
Clone the repository:
git clone https://github.com/commander-error404/chechen-tts.git
cd chechen-tts
Create a virtual environment (recommended):
python -m venv .venv
Activate it:
- Windows (PowerShell):
.\.venv\Scripts\Activate.ps1 - macOS / Linux:
source .venv/bin/activate
Install dependencies:
pip install -r requirements.txt
Run the script with the default example:
python main.py
Or pass your own Chechen text:
python main.py "Ассаламу алайкум, муха ду хьан гӀуллакхаш?"
On first run, the model (~150 MB) will be downloaded from Hugging Face and cached locally. Subsequent runs will start instantly.
The generated audio will be saved as output.wav in the project folder.
- Pronunciation of Chechen-specific sounds (кӀ, пӀ, тӀ, цӀ, чӀ, Ӏ, хь, гӀ, къ) is inconsistent
- Intonation can sound flat or unnatural
- Long sentences may lose rhythm — splitting text into shorter sentences helps
- Different orthography variants (Ӏ vs I vs 1) may produce different results
These issues are inherent to the baseline model and are the main motivation for the ongoing improved version.
- Working baseline (MMS-TTS-che)
- Collect and curate a Chechen speech dataset
- Fine-tune the model for improved naturalness
- Publish improved model weights
- Simple web demo
- Native Chechen speakers — feedback on pronunciation and naturalness
- Chechen texts — clean written material with clear licensing
- Chechen audio with transcripts — audiobooks, lectures, interviews with matching text
- ML engineers — training pipelines, evaluation, G2P
- Anyone — share the project in Chechen-speaking communities
Open an issue or reach out if you'd like to contribute.
MIT License. See LICENSE for details.
- Meta AI for the MMS project
- The Chechen-speaking community