Turkish speech to text: Measured accuracy and how to improve it

Author: Mert S. Kaplan · Last updated: November 3, 2026

Turkish speech to text works well today if you pick the model carefully: on clean recordings we measured about 90% of words correct with the best models, and about 62% with the smallest. LiveType is built around Turkish first, runs Whisper in your browser and lets you correct the text while it is written. Here is what we measured and what moves the numbers.

Why Turkish is hard for speech recognition

Turkish builds long words from suffixes, and a single word can carry what English needs a whole sentence for. That matters in two ways. One misheard suffix turns “geldi” (he came) into “geldik” (we came), and a word error rate counts the entire word as wrong. It also means negation can hang on one short sound, so “geldi” and “gelmedi” (did not come) are close enough to be dangerous in a quote.

Names and foreign terms add to this. A model may never have seen a local politician’s surname, and English terms inside a Turkish sentence sometimes get spelled by their sound. General models also train on far more English: in the same test, every Whisper size does better on English.

Try it on your own recording

No card needed. 30 minutes of transcription on the free plan.

Transcribe Turkish audioGo to the app

Measured accuracy by model

Model	Download	Words correct on Turkish, roughly	Runs on
M1 (Whisper Base)	165 MB	62%	Most devices, phones too
M2 (Whisper Small)	410 MB	81%	Devices with enough memory
M3 (Whisper Turbo, q4)	760 MB	90%	Graphics card
M4 (Whisper Turbo)	1.6 GB	91%	Graphics card

These come from short, clean recordings in the Turkish part of the FLEURS dataset (October 2026). A live broadcast or a crowded meeting will score lower, so use them to compare models, not to predict your result. The method and its limits are in our Whisper accuracy test.

Filler sounds

Spoken Turkish is full of “ııı”, “eee” and “hmm”. LiveType leaves these fillers out, so the text comes out readable. If hesitations matter to you, as they can in qualitative research, add them by hand from the audio.

Why names go wrong

Wrong names are the mistake people notice most. When the model has not met a surname in its training data, it picks the spelling closest to what it heard, and the same name can come out differently across one recording.

LiveType gives the model no list of names and no title: it writes each stretch from the sound alone. Editing while it transcribes pays off here, because you can fix a surname without waiting for the end. A correction does not teach the model, so if the same name comes out wrong again, fix it there too.

We cover other errors, such as numbers and invented sentences, under AI transcription accuracy.

Exporting Turkish transcripts

Download the text as Word (DOCX), ODT, Markdown, plain text, or timed SRT and VTT subtitles. Speaker names go into the document where the speaker changes, and Turkish letters stay as they are.

If the recording is sensitive, the default This device option keeps the audio in your browser. Private transcription explains what leaves your device and how to check. Other tools focus on Turkish too; see how Voiser and Transkriptor differ on their comparison pages.

What it does not do

It does not translate Turkish into another language.
We have not measured regional accents or very noisy rooms, where accuracy falls below these figures.
It does not proofread. It writes what the engine heard, and the final read is yours.

The free plan gives you 30 minutes in total, enough to try your own recording, and your first transcript shows the steps. Longer work is covered on the pricing page.

Mert S. Kaplan

Works in news, software and design; founder of LiveType. Worked as a reporter for many years. Has done SEO consulting, content editing and reporting for several online news sites, and has developed many pieces of software. Currently works in the press office of the Turkish Medical Association and provides SEO support to Elips Haber.