yagmurtuncer/turkish-asr-leaderboard
4
πΉπ· Turkish ASR Leaderboard π€
14 open Turkish speech-recognition models on one balanced test set, scored with WER, CER and KDHO, a root- and affix-aware error rate for Turkish. All values are error rates: lower is better.
π§ Try the models on your own audio: Turkish ASR Leaderboard Demo (Whisper, Voxtral and Voxtral + Turkish LoRA).
π Test set (TurkASR-Bench)
π₯ Download: datasets/yagmurtuncer/turkasr-bench (gated).
743 utterances from each of 7 sources: Common Voice 17, FLEURS, Khan Academy, MediaSpeech, VoxForge, YouTube and Real-TurnTurk (spontaneous dialogue).
- Fixed seed 42, equal weight per source.
- 5,200 utterances scored.
π€ Models
- Whisper: tiny, base, small, medium, large-v2, large-v3, large-v3-turbo
- Turkish Whisper fine-tunes:
turkmedstt/whisper-large-v3-turkish-general,selimc/whisper-large-v3-turbo-turkish - CTC models: XLS-R 300M (Common Voice 6 and 8 fine-tunes), MMS-1B-all (Turkish adapter)
- Multilingual models: Qwen3-ASR-1.7B, SeamlessM4T-v2-Large
π Metrics
- WER / CER: corpus-level, via
jiwer. - KDHO (KΓΆk ve Ek DuyarlΔ± Hata OranΔ±): a word-level edit distance.
- A substitution costs 0 if the words are the same.
- It costs λ·d_M if the words share a root and only the affix tags differ (λ = 0.5).
- It costs 1 if the roots differ, the negation differs (geldi β gelmedi) or a word cannot be analyzed.
- KDHO is never higher than WER.
- Normalization: NFC β Turkish lowercasing (IβΔ±, Δ°βi) β punctuation removed (word-internal apostrophes kept). The main setting also spells out digits in Turkish (48 β kΔ±rk sekiz).
- Significance: β marks a difference to the next model that is not significant (paired bootstrap, B = 2,000).
π§ Tabs
π« Contact
- Maintainer: yagmurtuncer
- Last updated: September 2026
- Questions or model submissions: open a discussion in this Space.
