01
Free to use
CC BY 4.0: personal or commercial, no fees, no sign-up. Just give credit.
FREE · OPEN · ON YOUR DEVICE
Drop a video or audio file and QiTV's open models write its Turkish subtitles on your own device. Nothing is uploaded.
Drop your video or audio file here
NOTHING IS UPLOADED · UP TO 30 MINUTES · .SRT / .VTT
You can choose a new file when this one is done.
Video or audio: MP4, WebM, MP3, M4A, WAV… Up to 30 minutes. The file never leaves your device.
The first time, the model downloads once (144 MB or 278 MB, shown before it starts) and your browser keeps it for your next visit. It cuts the speech at pauses and writes numbers as digits, as QiTV does: “saat on beş otuzda” → “saat 15.30'da”.
The subtitles appear under the video as they're written. Watch the video with them, then download .srt or .vtt.
We care about accessibility. Subtitles let deaf and hard-of-hearing people follow Turkish TV, films and talk, and they help everyone learning Turkish. Good Turkish captions shouldn't depend on a paid service or an internet connection.
01
CC BY 4.0: personal or commercial, no fees, no sign-up. Just give credit.
02
It runs offline on an ordinary CPU, on the viewer's own device. No audio leaves it.
03
The weights, the ONNX files, and the training and evaluation scripts are all public. Fine-tune it on your own data, shrink it, ship it in your app.
04
We test against every open Turkish model we could find and publish where others do better. If you build a better one, on ours or from scratch, tell us: we'd be proud to have helped.
MOONSHINE · TR + EN
The default Turkish model for QiTV's on-device AI subtitles. Turkish and English in one model; cased and punctuated.
Where others are better: on read speech and news, several open Turkish models are more accurate; mihu-v2 is the most accurate there.
On Hugging FaceMOONSHINE · TR + EN
For computers where the medium model is too slow, at half the size. It writes Turkish and English too; cased and punctuated.
Where others are better: on read speech and news, several open Turkish models are more accurate; mihu-v2 is the most accurate there. Our medium model is more accurate on every test.
On Hugging FaceFASTCONFORMER · TR ONLY
Built on NVIDIA FastConformer, Turkish only. It runs in your own app with sherpa-onnx or NeMo, and the NeMo file lets you keep training it. This page runs the Moonshine models.
Where others are better: on read speech and news, several open Turkish models are more accurate; mihu-v2 is the most accurate there, and on drama our Moonshine models are.
On Hugging FaceShare of words wrong (WER, %), lower is better; the lowest in each row is marked. The drama and song tests are private, the others public. Numbers and method from the model cards.
| Test | Medium | Small | Conformer | mihu-v2 | ufakzeka-kulak | balbal-v2 | zipformer-asr-turkish-v1 |
|---|---|---|---|---|---|---|---|
| TV channel A | 11.8 (lowest) | 14.3 | 16.5 | 19.2 | 24.8 | 26.2 | 22.4 |
| TV channel B | 15.0 (lowest) | 17.3 | 19.1 | 22.2 | 27.3 | 29.0 | 25.3 |
| TV channel C | 19.0 (lowest) | 22.0 | 23.1 | 25.1 | 31.5 | 33.8 | 28.4 |
| Sung lyrics | 38.9 (lowest) | 48.5 | 50.6 | 68.3 | 77.7 | 79.0 | 77.2 |
| FLEURS (read) | 11.7 | 13.6 | 11.4 | 7.9 (lowest) | 9.4 | 8.0 | 9.4 |
| Common Voice (read) | 10.0 | 13.1 | 11.0 | 6.0 (lowest) | 9.3 | 7.0 | 9.6 |
| MediaSpeech (news) | 12.2 | 13.4 | 12.7 | 10.8 (lowest) | 11.8 | 11.3 | 11.9 |
| Talk videos | 25.1 | 28.7 | 24.4 | 22.5 (lowest) | 24.8 | 24.4 | 23.1 |
Not only for QiTV: use them in your own app, your site, any project. CC BY 4.0; credit them as the model card asks and say what you changed.
The training, export and test scripts, the converter that writes numbers as digits among them. MIT.
QiTV is a free IPTV player for Windows, macOS and Linux. With these models it writes subtitles for what you watch, on your device.