Samples
Turkish speech to text: a real sample transcript
What the free preview returns for a short Turkish recording (Türkçe), with the audio, the sources and the mistakes left in.
Updated
Turkish (Türkçe), 28 seconds, taken 2026-10-27. Sentences read aloud by a volunteer. Output of the free preview with the language set to Turkish, not corrected.
The transcript, as the engine returned it
- Kaybettik ve verken dönmek zorundayız.
- Sırbistan Kara'da bu sefer altın madalleye uzanabilecek mi?
- Fabrika 2 merkezde faaliyet gösteriyor.
- Meclis kesintileri ekim ayında değerlendirecek.
- İkiliği der ayrıntılı bilgi vermedi.
- Kafe'de Miripan'ın telefonu çalıyor.
- Bunun süreci hızlandırması bekleniyor.
- Sorun tek bir sporla sınırlı değil.
- Peki kostün icada ha ne kadar dayanabilir?
Download this transcript as SRT (written by the same exporter the app uses; the other formats are TXT, timestamped TXT, VTT, Markdown, JSON and CSV).
Against the reference sentences
Word error rate 30.0% (15 edits over 50 words), character error rate 6.2%; normalisation: Turkish (explained on the samples page).
The sentences the speaker was asked to read:
- Kaybettik ve eve erken dönmek zorundayız.
- Sırbistan-Karadağ bu sefer altın madalyaya uzanabilecek mi?
- Fabrika iki merkezde faaliyet gösteriyor.
- Meclis kesintileri Ekim ayında değerlendirecek.
- İki lider ayrıntılı bilgi vermedi.
- Kafede Mripa'nın telefonu çalıyor.
- Bunun süreci hızlandırması bekleniyor.
- Sorun tek bir sporla sınırlı değil.
- Peki Kostunica daha ne kadar dayanabilir?
| In the reference sentence | In the transcript |
|---|---|
| eve | (missing) |
| erken | verken |
| (nothing: an extra word) | kara |
| karadağ | da |
| madalyaya | madalleye |
| iki | 2 |
| iki | ikiliği |
| lider | der |
| (nothing: an extra word) | kafe |
| kafede | de |
| mripa | miripan |
| nın | ın |
| (nothing: an extra word) | kostün |
| kostunica | icada |
| daha | ha |
Where the recording comes from
- Source: Mozilla Common Voice 22.0, test split (CC0 1.0 public domain dedication, https://commonvoice.mozilla.org/en/datasets, https://creativecommons.org/publicdomain/zero/1.0/; copy used: https://huggingface.co/datasets/fsicoli/common_voice_22_0, whose card also says cc0-1.0).
- Licence: CC0 1.0 (creativecommons.org/publicdomain/zero/1.0/).
- Which recording: One volunteer speaker (male, twenties): the speaker with the most clips in the part of the test archive we read, their clips in the test file's order that are 2.5 to 7 seconds long and have no down-votes, until about 24 seconds were collected. Chosen by that rule, not by how well the engine did.
- Clip ids: common_voice_tr_17525375, common_voice_tr_17525376, common_voice_tr_17525379, common_voice_tr_17525388, common_voice_tr_17525405, common_voice_tr_17525407, common_voice_tr_17525425, common_voice_tr_17525430, common_voice_tr_17525431.
How every sample was made, step by step, is on the samples page. To try your own Turkish recording, use the Turkish transcription page.
Questions
What did the engine get wrong in this Turkish clip?
Can I use the Turkish recording?
Try it with your own file
Upload a recording and read the transcript.