voice-models
Training recipes, configs, dataset metadata and benchmarks for the Kurmanji (Northern Kurdish, kmr) speech models that dengjen runs.
The voices are built for blind people using NVDA, often at two to three times normal speed, so clarity and low latency come before expressiveness.
The repository holds code and configuration only. Model weights live on Hugging Face and audio lives outside git; see Where weights and audio live.
Current models
| Voice | Engine | Quality | Licence | Weights |
|---|---|---|---|---|
berfin-renas |
Piper |
medium |
The berfin-renas voice is trained on recordings provided by Zarok TV in partnership with Zirek.
Roadmap
-
Piper
berfin-renasatlowandhighbeside the existingmedium. -
Kokoro and other low-latency engines trained on the same data and compared against Piper on the shared benchmark.
-
More voices from new recordings of the recording script. To take part, see Record your voice.
-
Speech-to-text: a Whisper fine-tune for Kurmanji.
Repository layout
tts/ Text-to-speech, one folder per engine
piper/ Piper (VITS) recipes, low / medium / high quality
voices/<voice>/ Per-voice config: speaker, dataset, HF repo, per-quality settings
kokoro/ Kokoro (StyleTTS2-based) experiments
stt/ Speech-to-text (planned)
whisper/ Whisper fine-tune
data/ Dataset docs and manifests (text only, no audio)
recording-script/ Kurmanji sentences for speakers to record
manifests/ CSV/JSONL manifests pointing at audio stored elsewhere
eval/ Shared latency and quality benchmarks for every model
sentences/ Fixed test sentences, kept out of every training set
results/ Benchmark summaries, one file per model and run
docs/ This documentation
Each engine folder has the same shape: a README.md with how to train and export, and voices/<voice-id>/ with one config per voice.
Adding a voice means adding a folder, not changing engine code.
