Software that reads and speaks the languages large vendors don't get around to.
Two strands, and they meet: screen-reader support for under-served languages, and the local neural speech synthesis that makes a screen reader usable in the first place. Projects, releases and docs: zirekhq.github.io.
- dengjen-nvda (overview) — NVDA add-on for fast, fully local neural text-to-speech using Piper voices. No cloud service, no account, no network. In the official NVDA Add-on Store. Maintenance continuation of mush42/sonata-nvda, renamed from Sonata Neural Voices at the original author's request as a condition of the store listing.
- nvda-addon-testkit (overview) — end-to-end testing for NVDA add-ons against a real NVDA in CI, covering the seams an add-on actually needs: speech, braille, keys, config, log.
The stack underneath the add-on, all continuations of Musharraf Omer's work:
- dengjen (overview) — cross-platform inference engine for neural speech models (Rust): text to speech with pluggable backends (Piper, Kokoro) and offline speech recognition. Includes the sentence segmentation that used to be a separate project (tqsm), tuned for speed over linguistic perfection — splitting text correctly is most of what makes synthesised speech sound unhurried.
- dengjen-piper-rs (overview) / dengjen-tts-go (overview) — the Piper backend crate, and Go bindings for dengjen (speech synthesis and recognition) with prebuilt native binaries.
- dengjen-tashkeel (overview) — diacritic restoration for Arabic. Arabic is normally written without the short vowels a synthesiser needs, so they have to be inferred before anything can be spoken.
Kurmanji phoneme support is contributed upstream to espeak-ng rather than kept here, so it reaches every project that uses it.
- dengjen-werger (overview) — commitment and peer-review layer for sustained Kurmanji translation contribution, sitting in front of Crowdin.
- nvdaku — Kurdish (Kurmanji) localisation for NVDA from 2018: a partial interface translation, character descriptions and a symbol dictionary. No longer maintained.
Kurmanji is badly served by assistive technology, and most of the gap needs native speakers rather than programmers:
- Voice recordings for text-to-speech training.
Sorani and Zazaki speakers are just as welcome; the work is Kurmanji-first only because that is where it started.
On the synthesis side the shortage is different:
- Windows testers for dengjen-nvda, especially against current NVDA releases.
- Co-maintainers. The add-on has one maintainer, which is one too few.
Open an issue on the relevant repository — for anything that sounds wrong, include the exact text and what it should sound like. See CONTRIBUTING.md; you do not need to write code to help.
Zirek is maintained in spare time and paid for out of pocket. Donations through Open Collective pay for compute, model training, dataset quality, testing hardware and contributor bounties, and every expense is public.