Note
From a dictionary to a voice assistant: the audio companion to the first article
Where the first stage of the road ended up: speech recognition from 34.6 % down to 17.0 % word errors, a whole novel — 16 chapters, 15,444 words, 1 hour 58 minutes — read by a synthetic voice, all of it on one consumer GPU. With samples you can listen to.
The first article of the series is out on ana-yurt.com, and it is about where all of this started: Crimean Tatar has almost no language technology at all. Not «little» — almost none. No usable machine translation, no speech recognition, no synthesis to simply listen to a text. Meanwhile the language sits on the endangered lists. This page is the audio companion to that article: what is described there in words can be heard here.
The plan that grew out of it still sounds presumptuous: walk the whole road, from a plain dictionary to a voice assistant you can actually talk to in Crimean Tatar. The dictionary and the Cyrillic-to-Latin conversion already work. Speech turned out to be the next stage, and the hardest one — and it came before translation for two reasons. The people whose living speech can still be recorded are not getting younger; that is the one part of the job that cannot wait. And recognition and synthesis pull each other up: old recordings become text, text with audio becomes training material for a voice, the voice makes audiobooks, people listen, and there is more speech to learn from.
Where this stage ended up. Recognition: the model I started with got 34.6 % of words wrong; after fine-tuning, 20.1 %; after carefully tuning how the model picks words while decoding, 17.0 % — half the errors it began with, and the whole add-on I trained weighs 126 MB. Synthesis: the voice read an entire book — Şamil Alâdin’s novel «Merdiven», not an excerpt and not a single showcase chapter, but all 16 chapters — 15,444 words, 1 hour 58 minutes, with word-level timings verified in every chapter. All of it on one consumer GPU, in the evenings and at night, with no recording studio, no hired narrators and no server farm.
The sound /q/ deserves its own note. An off-the-shelf Turkish voice read qara as «kara» — Turkish
simply does not have our /q/. What fixed it was not more data but a different foundation: a
multilingual base trained on Kazakh, Uyghur and Bashkir, languages where the sound exists and has a
letter of its own, plus rewriting our къ, гъ, нъ into the ҡ, ғ, ң it already knows. The pronunciation
was right before any fine-tuning at all. You can hear it below, in words we say every day.
Audio
Everything below is synthesis unless stated otherwise; the voice is Sevil, cleared for demo use.
Merdiven, chapter fragment
Words with q, ğ, ñ
Human and model
The same three sentences, first read by a person, then by the model. Two separate players; listen top to bottom.
Human
Model
How it started and how it is going
The first model I put out publicly was SpeechT5, December 2025: Crimean Tatar Latin was rewritten into English phonetics (ğ→gh, ş→sh, ñ→ng) and the audio came out at 16 kHz. The two clips do not read the same text: that first model was only ever probed word by word, because words were as far as it went.
First attempt, December 2025
Current model
Read the full story
- The full article on ana-yurt.com (in Russian)
- Medium (EN) — My Language Is Listed as Endangered. So I Built It a Voice.
- dev.to (EN) — What I learned building speech tech for a language with zero datasets
- Hugging Face blog (EN) — Speech tech for a zero-dataset language: what actually consumed the time
- Хабр (RU)
- LinkedIn (EN) — Giving an endangered language a voice, on one GPU