DNShitobu/dagbani-tts

Single-speaker Dagbani (dag) text-to-speech: VITS/MMS-TTS fine-tuning pipeline, data inventory, and Colab notebook. # Dagbani Text-to-Speech — Build Guide & Pipeline A complete, Colab-ready pipeline to build a high-quality **single-speaker Dagbani (`dag`) TTS** voice by fine-tuning a VITS / MMS-TTS model. This repo contains: | File | Purpose | |---|---| | `DATA_INVENTORY.md` | Every usable Dagbani dataset + model found, tiered, with licenses. **Read this first.** | | `Dagbani_TTS_Finetune_Colab.ipynb` | One-click-ish Colab notebook: install → data → train → listen. | | `dagbani_text_normalize.py` | Dagbani text normalization (used at train **and** inference time). | | `prepare_dagbani_dataset.py` | Download → clean → filter → split into a training `DatasetDict`. | | `finetune_dagbani_vits.json` | Training config for `ylacombe/finetune-hf-vits`. | | `synthesize_dagbani.py` | Inference: text → WAV with your fine-tuned model. | --- ## TL;DR — the strategy * **Meta MMS-TTS has no Dagbani checkpoint.** But a Dagbani VITS model already exists (`FarmerlineML/dagbani_tts-2025_v2`, *the same architecture as MMS-TTS*). So we **warm-start from that** and fine-tune on a clean single-speaker corpus. * **Best clean-voice data** = the **Dagbani audio Bible** (one professional narrator, 24 kHz, already force-aligned & VAD-trimmed): `ghananlpcommunity/dagbani-bible-audio-text-tts` (~53k segments, CC-BY-4.0). * **Avoid** the "navigation corpus" *as a voice target* — it is synthetic (TTS-generated) and would teach the model its artifacts. Fine for text/ASR. * **Don't** train a model you intend to release on **UGSpeechData** — its CC-BY-NC-ND license forbids distributing derivatives. --- ## Step-by-step (matches the notebook) ### 0. Runtime In Colab: **Runtime → Change runtime type → GPU** (T4 is enough; L4/A100 faster). ### 1. Install the finetuner + build the alignment kernel ```bash git clone github.com cd finetune-hf-vits pip install -r requirements.txt pip install datasets[audio] librosa soundfile accelerate # Build the Cython monotonic-alignment sear …