Bitig Large 1.0 is a Turkish speech recognition model for products: calls, meetings and media. It is whisper-large-v3 fine-tuned only on Turkish speech whose licences allow commercial use, and it writes cased, punctuated text. It is released under Apache-2.0.
| Developer | Cagan Oflazoglu |
| Model type | Whisper encoder-decoder, 32 + 32 layers, 1.54B parameters |
| Language | Turkish |
| Fine-tuned from | openai/whisper-large-v3 |
| Formats | transformers (float16) and CTranslate2 for faster-whisper |
| Licence | Apache-2.0 (LICENSE, NOTICE) |
| Version | 1.0, tag c1.0, November 2026 |
| Research sibling | Kashgari Large 1.0, trained on more data, research use only |
Word error rate (WER, %) after the normalizer tr-norm-v1 (lowercase, punctuation removed, numbers written as
words), decoded with CTranslate2 in float16, beam 5, no timestamps.
I scored the test sets once, on 13 November 2026, with the model, the decoding settings and the normalizer fixed before the run, and scored whisper-large-v3 again in the same run. The difference row is Bitig's WER minus whisper-large-v3's, with a 95 % interval from a paired bootstrap (10,000 resamples; FLEURS resampled by sentence and parliament by session).
| FLEURS tr test (743 clips) | Common Voice 17 tr test (2,000) | Parliament, 10 sessions (1,000) | MediaSpeech tr (1,000) | |
|---|---|---|---|---|
| Bitig Large 1.0 | 4.69 | 6.09 | 9.48 | 10.16 |
| whisper-large-v3 | 6.05 | 7.90 | 9.59 | 11.21 |
| Difference | -1.37 [-1.85, -0.91] | -1.81 [-2.30, -1.32] | -0.12 [-0.42, +0.11] | -1.05 [-1.60, -0.49] |
On parliament speech, which Bitig never trained on, it is level with whisper-large-v3. An audio fingerprint check of the training audio against every test set found 4 of the 1,000 MediaSpeech test clips inside audio Bitig trained on: 3 in Voice of America broadcasts and 1 in the ISSAI corpus. Without them Bitig scores 10.18 and whisper-large-v3 11.23, so the figure above stands. The check finds about 6 in 10 degraded copies, so it can miss some, and the Common Voice 17 and FLEURS train audio is not checked yet.
These boards ran the model themselves; the figures are theirs, as listed on 19 November 2026.
| Board | What it tests | Bitig Large 1.0 | whisper-large-v3 | Place |
|---|---|---|---|---|
| TurkASR-Bench | 5,200 utterances from seven sources, numbers spelled out | 12.04 | 12.83 | 1st of 26 |
| COSMOS Speech Benchmark | 200 real phone calls at 8 kHz | 24.45 | 28.03 | 7th of 13 and the second-best open model, behind Kashgari Large 1.0 (22.45) |
On TurkASR-Bench, 83 of the 743 YouTube rows share text with Bitig's training data and two MediaSpeech clips are in its training audio. The board's clean score leaves out the whole YouTube source and those clips: 12.10 on 4,455 utterances.
| FLEURS tr (338 clips) | Common Voice 17 tr (2,000) | Parliament (1,000) | |
|---|---|---|---|
| Bitig Large 1.0 | 5.94 | 5.15 | 8.57 |
| Kashgari Large 1.0 | 5.27 | 4.56 | 5.67 |
| whisper-large-v3 | 7.48 | 6.71 | 8.72 |
Parliament is five whole sessions of the Turkish parliament, from WorldSpeech, that neither model trained on.
Phone and room audio, FLEURS tr dev through simulated channels:
| Channel | whisper-large-v3 | Bitig Large 1.0 |
|---|---|---|
| Landline (G.711 A-law, 8 kHz) | 10.57 | 7.77 |
| Mobile (AMR narrowband, 7.4 kbit/s) | 11.60 | 8.96 |
| Internet call (Opus, 12 kbit/s) | 8.36 | 6.46 |
| Far room (reverberation, pink noise at 10 to 15 dB SNR) | 25.77 | 19.30 |
With faster-whisper and the CTranslate2 build in ct2/:
from faster_whisper import WhisperModel
model = WhisperModel("ct2", device="cuda", compute_type="float16") # "int8" on a CPU
segments, _ = model.transcribe("audio.wav", language="tr", beam_size=5, without_timestamps=True)
print(" ".join(s.text.strip() for s in segments))
With transformers:
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="turkish-sa-asr/bitig-large-v1.0", dtype="float16", device="cuda")
print(asr("audio.wav", generate_kwargs={"language": "tr", "task": "transcribe", "num_beams": 5})["text"])
Every score on this card comes from the CTranslate2 build. On the dev sets, transformers with the settings above
scored 0.20 points higher WER on FLEURS (95% interval [+0.04, +0.39]) and 0.17 lower on Common Voice 17
([−0.38, +0.02]) (results/eval/backend_equiv.json).
The model was trained without timestamp tokens. For recordings longer than 30 s, use voice activity detection
(eval/transcribe.py --long does).
eval/ reproduces the numbers above:
pip install -r eval/requirements.txt
python eval/smoke.py . # FLEURS tr dev: prints WER 5.94 on a CUDA GPU in float16
python eval/transcribe.py . clips/ hyps.jsonl # decodes a folder with the settings above
python eval/score.py refs.jsonl hyps.jsonl [--raw] # corpus WER and CER after tr-norm-v1
For a fair comparison, score with eval/normalize.py, or report raw WER next to your own normalizer: Whisper's
English normalizer maps I to i instead of ı. Please report the tag (c1.0) or the commit hash with your scores.
Bitig trained on the train splits of FLEURS, Common Voice 17 and the ISSAI Turkish Speech Corpus, never on their dev or test splits, and never on WorldSpeech or the MediaSpeech dataset. Some MediaSpeech recordings also appear in Voice of America broadcasts and in ISSAI, though: 4 of the 1,000 test clips above were found in Bitig's training audio. Three things to check before scoring:
training/common_voice_clips.txt, and
check the test split against CV17's train split, since releases can re-split clips.998a8657-04ec-4c69-9328-c4bec9ab57bb, 42afea1f-7828-498d-88b7-9415ed2d4d0f,
9d5eeb73-73fe-4e8a-b53e-718325a91de1, f8e65143-55d7-4f2b-ad5f-26ed510d2485 (.flac in ymoslem/MediaSpeech).| Source | Windows | Licence |
|---|---|---|
| ISSAI Turkish Speech Corpus, train (transcripts repaired with Whisper's punctuation and casing) | 26,701 | MIT |
| Voice of America Türkçe, staff-produced items, labelled by whisper-large-v3 (296.62 h) | 41,413 | US public domain |
| Common Voice 27 tr, clips not in CV17 (41.85 h) | 5,418 | CC0-1.0 |
| Common Voice 17 tr, train, seen twice per epoch | 2,670 | CC0-1.0 |
| FLEURS tr, train, seen twice per epoch | 1,036 | CC BY 4.0 |
A window is up to 30 s of audio. VOA and Common Voice 27 sentences sharing an 8-word sequence with a dev or test text were dropped. Training ran 3 epochs (3,789 steps at batch 64, learning rate 1e-5) on a Kaggle TPU v5e-8, with room and noise augmentation from OpenSLR 28.
Apache-2.0 (LICENSE, NOTICE), released 19 Nov 2026. Common Voice 27's terms ask that outputs made with
its data are marked as machine-generated, so please mark the model's transcripts that way.
@misc{oflazoglu2026bitig,
author = {Oflazoglu, Cagan},
title = {Bitig Large 1.0},
year = {2026},
url = {https://huggingface.co/turkish-sa-asr/bitig-large-v1.0}
}
Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., Rivera, C., & Bapna, A. (2023). FLEURS: Few-shot learning evaluation of universal representations of speech. 2022 IEEE Spoken Language Technology Workshop (SLT), 798-805.
Mussakhojayeva, S., Dauletbek, K., Yeshpanov, R., & Varol, H. A. (2023). Multilingual speech recognition for Turkic languages. Information, 14(2), 74.
Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2023). Robust speech recognition via large-scale weak supervision. Proceedings of the 40th International Conference on Machine Learning, 28492-28518.
ct2/model.bin 3.1 GBmodel.safetensors 3.1 GBggml/ggml-bitig-large-v1.0-q8_0.bin 1.7 GBggml/ggml-bitig-large-v1.0-q5_0.bin 1.1 GBopenai/whisper-large-v3