VibeusTurkish speech recognition on a CPU: the developer of balbal-v2... - Vibeus
Turkish speech recognition on a CPU: the developer of balbal-v2 published benchmark results for two CTC + KenLM models. The 313M-parameter Manas recognizes Turkish and English and is reported to run at about 123× real time on eight threads. The 45M-parameter model handles Turkish only and was reported at about 236× real time. The smaller model also ran faster than real time on one core of an older Celeron running Windows 7. The developer puts training costs for both below $250.
On Mu-Bench Turkish, a set of 846 customer service calls, the developer reports 8.20% word error rate and 14.89% sigWER for Manas using the official harness. Google Chirp 3 scored 8.03% and 14.30%, and the 45M balbal-v2 scored 9.46% and 18.44%. On Turkish FLEURS, Manas scored 4.06% WER versus Whisper large-v3-turbo’s 4.07% and was reported to run about 75× faster on the same CPU. The English result was 8.67% average WER across the Open ASR leaderboard sets.
The architecture uses product-key memory inspired by engrams and a refiner that runs three times before output. In the discussion, the developer explained that the microphone mode processes overlapping audio windows because the model attends to a whole clip. A separate team, LocalDoc, reported 10.98% WER on Azerbaijani FLEURS after training a model on the Manas architecture.
The package can transcribe a WAV from the command line, run a live microphone mode or serve an OpenAI-compatible endpoint. The code is Apache-2.0. The model weights have separate terms covering research, noncommercial use and eligible small businesses, with restrictions on resale as an ASR product or API.
