Tachelhit · Taclḥiyt · ⵜⴰⵛⵍⵃⵉⵜ
Speech to text, in Tachelhit.
Record a short phrase or upload an audio file. The model transcribes it and restores Tachelhit word boundaries with the Tussna lexicon.
Recorder
or
Noureddine corpus · MIT
Raw model output (before word segmentation)
About this demo
- Model: OmniASR CTC 300M, fine-tuned on 2.45 h of Tachelhit speech (MTNRA + Noureddine).
- Word boundaries: restored with the Tussna Tachelhit spellchecker lexicon (197k forms).
- Limitations: research demo, not production. The test set is small and partly text-overlapping; expect errors, especially on names, code-switching, and noisy audio.
- Privacy: audio is processed on the server and deleted immediately. Nothing is stored or logged.