DiaCTC — Arabic Diacritization from Speech

Constrained CTC / WFST decoding for efficient diacritic restoration. Model: rufaelfekadu/diactc-wav2vec2-clartts

Provide the undiacritized Arabic transcript of the audio and upload or record the matching speech. DiaCTC restores diacritics using constrained WFST decoding over a character-level diacritization lattice.

Tip: The transcript should match what is spoken (e.g. ASR output without tashkeel).
This demo runs on CPU — keep clips under ~30s; inference may take a minute or two.

Undiacritized transcript

Speech audio

Decoding method

Constrained WFST decoding

Diacritized output

Example

Undiacritized transcript	Speech audio	Decoding method	Constrained WFST decoding

Built with DiaCTC · Accepted at Interspeech 2026