NartOS/mms-1b-abaza-asr

Abaza speech recognition (MMS-1B adapter)

An open speech-to-text model for Abaza (абаза бызшва, abq), one of the smallest languages of the Caucasus. Speak Abaza — get Cyrillic text. To our knowledge, the first open speech recognition model for the language.

Results

Official Mozilla Common Voice 27.0 test set: 707 utterances from 14 speakers the model has never heard (test and dev speakers are excluded from training).

Metric	Value
Character error rate (CER)	6.7%
Word error rate (WER)	42.3%

Abaza words are long, so one wrong letter or space makes the whole word wrong — CER is the more informative metric.

How it was trained

• Base: facebook/mms-1b-all (965M parameters), frozen.
• Trained: a language adapter (~2.2M parameters) and the output layer only.
• Initialized from the MMS Abkhaz adapter — Abkhaz is Abaza's closest relative.
• Data: Mozilla Common Voice 27.0, Abaza, CC0 — only 2.7 hours from 3 speakers. ~1,000 steps, effective batch 32, fp16, on free Kaggle T4 GPUs.
• Text normalization: all palochka variants (I, l, 1, |, і) unified to ӏ; lowercase; no punctuation.

Usage

import librosa, torch
from transformers import AutoProcessor, Wav2Vec2ForCTC

repo = "NartOS/mms-1b-abaza-asr"
processor = AutoProcessor.from_pretrained(repo)
model = Wav2Vec2ForCTC.from_pretrained(repo).eval()

audio, _ = librosa.load("speech.wav", sr=16000, mono=True)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
with torch.inference_mode():
    ids = torch.argmax(model(**inputs).logits, dim=-1)
print(processor.batch_decode(ids)[0])

The repository also contains adapter.abq.safetensors (~9 MB) to swap onto a shared facebook/mms-1b-all instance.

Limitations

• Cyrillic output only, lowercase, no punctuation.
• Trained on just three voices: expect lower quality on other speaking styles, dialects and noisy audio.
• No language model.

License

CC BY-NC 4.0, inherited from the base MMS model: free for research and non-commercial use.

Authors

Magamed Addaj and Bislan Beretar — NartOS.

Versions

• v0.1 — first public release.

Citation

@misc{nartos_abaza_asr_2026,
  title  = {Abaza speech recognition: MMS-1B adapter},
  author = {Addaj, Magamed and Beretar, Bislan},
  year   = {2026},
  url    = {https://huggingface.co/NartOS/mms-1b-abaza-asr}
}