Open-Source Turkish ASR "balbal-v2": 236x Real-Time on CPU, Second Only to Chirp 3 on Mu-Bench
Shadomia · reddit · 2026-11-03
The author released balbal-v2, a Turkish speech recognition model—two versions, both CTC + KenLM and designed for CPU: the 313M balbal-v2-manas (Turkish/English bilingual, 123x real-time on 8 threads) and the 45M balbal-v2 (Turkish-only, 236x real-time, running faster than real-time even on a single core of an old Celeron POS machine). On the Mu-Bench TR leaderboard of 846 real customer-service calls, manas ranks second with WER 8.20 / sigWER 14.89, behind only Google Chirp 3 (8.03/14.30), beating Gemini 3.5 Transcribe, ElevenLabs Scribe v2, Deepgram Nova-3, and Azure, while whisper-large-v3-turbo lands last (24.65). On FLEURS Turkish, manas essentially ties whisper-large-v3-turbo (4.06 vs 4.07) but is 75x faster on the same CPU. The architecture uses an engram-inspired product-key-memory plus a three-pass refiner. Another team has already trained an Azerbaijani model with this architecture. Open-sourced under Apache-2.0, installable via pip install balbal, with a live microphone mode and an OpenAI-compatible server; weights are free for research/non-commercial/small businesses, but may not be resold as ASR products or APIs.
More from Models
- Dev compares astra and sol to Opus 5.5: rigid on open-ended tasks — haider1 · 2026-11-03
- Open-source models now beat every 'frontier' model from just 11 months ago — Hesamation · 2026-11-03
- Sebastian Raschka Maps October's 'Open October': 10 Open LLM Architectures in One Month — rasbt · 2026-11-03
- First hands-on: Meta's Muse Wonder 1.0 vs DeepSeek V4.1 Flash in Codex CLI — Richidm · 2026-11-03
- "Claude Code has been terrible today" — developer smells a new Opus launch coming — hermanschutte · 2026-11-03
- Mistral releases open-weight ML4 with frontier code-agent and cyber capabilities — _AndrewZhao · 2026-11-03