01
Sevil voice dataset
1,566 corrected recordings with transcripts, split into training, validation and test sets. Licence: CC BY 4.0.
About the project
Major voice services barely support Crimean Tatar. This project collects real voices and builds the first open models: one reads text aloud, the other turns speech into text.
It is built on a dataset of 1,566 recordings of Sevil’s voice, about 2.6 hours of speech. Only 415 files in the original set were usable; now all of them are fixed: noise, volume, pauses and transcripts. SpeechT5 (text-to-speech) and Whisper Large v3 (speech recognition, word error rate 13.7%) were fine-tuned on these recordings.
A whole novel has already been read with the synthesized voice: Shamil Alyadin’s “Merdiven”, 16 chapters, almost two hours. All models and data are open (CC BY 4.0 and Apache 2.0) and can be built into your own apps.
- 1 566
- voice recordings
- ≈2,6
- hours of speech
- 13,7 %
- recognition errors
- 16
- chapters of a novel read aloud
What it does
1
Sevil voice
A dataset of corrected voice recordings.
2
Text-to-speech
The first model that reads text aloud.
3
Speech recognition
The first model that turns speech into text.
Inside
02
Speech recognition
Whisper Large v3 fine-tuned on Crimean Tatar. It handles both Cyrillic and Latin script; word error rate is 13.7%.
Made by
Who makes it: Servin Osmanov (Ana Yurt).
Found a mistake? Write to us and we will fix it in the next update.
selam@ocaq.ru More in the house
