Madhav-Bhanushali/kreol
Domaine:
natural language processingType de record:
software
Créateur:
Mad
Hôte:
# Kreol — Mauritian Creole (Morisyen) TTS fine-tune pipeline
Fine-tunes **F5-TTS** on Mauritian Creole (`mfe`) using the NTKM2009 Bible audio
(see `morisyen-tts-project-status.md` for full project status, rationale, and license notes).
**⚠️ Licensing / status before you invest time:**
- F5-TTS's pretrained checkpoint is **CC-BY-NC** — any fine-tune inherits non-commercial status.
- The Bible text (© Bible Society of Mauritius, 2009) is **not** public domain. Confirm usage
rights (DBL entry app.thedigitalbiblelibrary.…) before
building on it. This pipeline is currently a **research/feasibility build**.
- The downloaded audio is the FCBH **Drama** recording (multi-speaker + background music), not a
single narrator. This is the only FCBH audio for this translation; Meta's MMS used the same
recording family for `mms-tts-mfe` after removing music and hard quality filtering.
## Hardware target
A single **NVIDIA RTX PRO 6000 Blackwell (96 GB)** GPU (sm_120). bf16 mixed precision, one process.
Blackwell needs CUDA 12.8+ — install `torch==2.8.0+cu128 torchaudio==2.8.0+cu128` (see F5-TTS README).
## Setup
```bash
# 1. F5-TTS (editable install required for training)
git clone github.com
cd F5-TTS && pip install -e .
conda install ffmpeg # or install ffmpeg some other way; required
# 2. This pipeline's deps
pip install -r requirements.txt
# 3. Forced-aligner
pip install git+github.com
```
## Pipeline
Run order (steps 2 and 3 are only needed because the audio is a drama recording):
| Step | Command | Notes |
|---|---|---|
| 0. Fetch text | `python scripts/00_fetch_verses.py --audio-dir "MFEBSMN2DA/Mauritian Kreol_mfe_BSM_NT_Drama"` | writes `data/verses.json` from bible.com (NTKM2009) |
| 1. Audio → WAV | `python scripts/01_convert_audio.py "MFEBSMN2DA/Mauritian Kreol_mfe_BSM_NT_Drama"` | MP3 → 24 kHz mono WAV |
| 2. Music/SFX removal | `python scripts/02_rem …