The post has been translated automatically. Original language: Russian
The copyright holder is "Public Health services" Sovetkhan Shyngys wants to share an open-source project: Kazakh Whisper Large-v3 Turbo (a model for recognizing Kazakh speech published on Hugging Face).
The model is based on Whisper Large-v3 Turbo and has been fine-tuned to 1,500+ hours of Kazakh speech from open datasets.
According to benchmark comparisons, the model shows 11.80% WER and 4.98% CER on FLEURS Kazakh.
Among the open-source Kazakh ASR models that were tested, it showed the best quality and surpassed baselines like wav2vec2-large-xlsr-kazakh and the basic Whisper Large-v3 Turbo.
Model: https://huggingface.co/shyngys879/kazakh-whisper-large-v3-turbo
The model is completely open and free. I will be glad if it turns out to be useful to developers, startups and companies that work with Kazakh audio, transcription, call center audio, voice assistants, subtitles or speech analytics.
Предлагаем вашему вниманию уникальный open-source проект: Kazakh Whisper Large-v3 Turbo (модель для распознавания казахской речи, опубликованную на Hugging Face).
- Модель основана на Whisper Large-v3 Turbo и была fine-tuned на 1,500+ часах казахской речи из открытых датасетов.
- По benchmark-сравнениям, модель показывает 11.80% WER и 4.98% CER на FLEURS Kazakh.
- Среди open-source Kazakh ASR моделей, которые были протестированы, она показала лучшее качество и превзошла baselines вроде wav2vec2-large-xlsr-kazakh и базового Whisper Large-v3 Turbo.
- Model: https://huggingface.co/shyngys879/kazakh-whisper-large-v3-turbo
Модель полностью открытая и бесплатная.
Предназначено: разработчикам, стартапам и компаниям, которые работают с казахским audio, transcription, call-center audio, voice assistants, subtitles или speech analytics.
Разработчик Советхан Шынғыс - выпускник РФМШ 2026, владелец авторского права ИС "Public Health services".
Share
