mms-tts-kin-onnx

MMS-TTS Kinyarwanda — ONNX export

ONNX export of facebook/mms-tts-kin (Meta MMS, VITS) for use with onnxruntime. Exported with optimum.

License of the underlying model: CC-BY-NC 4.0 (non-commercial).

Getting natural, professional-sounding output

The model is char-level: it garbles or drops digits and symbols. For speech that sounds like a real Rwandan announcer, normalize the text before tokenizing:

1. Numbers → Kinyarwanda words: 2500 → ibihumbi bibiri na magana atanu; prices 2500 RWF → amafaranga ibihumbi bibiri na magana atanu; percentages 15% → cumi na gatanu ku ijana; clock times 14:30 → saa cumi na kane n'iminota mirongo itatu; phone numbers digit-by-digit.
2. Fold characters outside the vocab: é→e, ç→c, q→k, x→gs, etc. (vocab is a–z minus q/x, plus ' - and digits).
3. Chunk on sentence/comma boundaries and join with 130–280 ms of silence — VITS run-on delivery becomes a calm newsreader rhythm.
4. Steadier delivery: speaking_rate ≈ 0.92, noise_scale ≈ 0.6, noise_scale_duration ≈ 0.7, and a fixed seed for a consistent voice.

A hosted implementation of all of this (REST POST /tts → WAV) runs at gloire2/languagetool — see its app.py for the reference normalizer (normalize_kinyarwanda).

Tokenization (this ONNX export)

Lowercase → per-char vocab lookup → interleave pad id 0 between ids ([0, id, 0, id, …, 0]), inputs input_ids + attention_mask (int64), output waveform (float32 @ 16 kHz).