AI Briefingbalbal-v2 Turkish ASR models achieve 8.20 WER on Mu-Bench TR, ranking second behind Google Chirp 3
·2026.11.03 22:57
Key point
The larger balbal-v2-manas model, a CPU-optimized Turkish ASR, ranks second on Mu-Bench TR behind Google Chirp 3, outperforming Gemini 3.5 Transcribe and ElevenLabs Scribe v2.
Details
- balbal-v2-manas (313M parameters) supports Turkish and English, while balbal-v2 (45M parameters) supports Turkish only; both are optimized for CPU inference.
- On the Mu-Bench TR benchmark (846 real customer service calls), balbal-v2-manas ranks second with 8.20 WER, trailing only Google Chirp 3 (8.03 WER) and outperforming Gemini 3.5 Transcribe (8.68 WER), ElevenLabs Scribe v2 (9.43 WER), and whisper-large-v3-turbo (24.65 WER). The smaller balbal-v2 achieves 9.46 WER.
- The models utilize a unique architecture featuring product-key-memory inspired by engrams and a refiner that loops three times before output; balbal-v2-manas runs at approximately 123x real-time speed on 8 threads, while the smaller balbal-v2 runs at 236x real-time speed.
- Code is released under Apache-2.0, and weights are free for research, non-commercial use, and small businesses, with a restriction against reselling the models as an ASR product or API.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.