Request access to KSTU ASR (manual review)
Access is reviewed manually and granted at the authors' discretion. This model is available to researchers and practitioners both inside and outside KSTU. Please tell us who you are and how you intend to use the model. Provide accurate information; incomplete or unclear requests may be declined.
Access is granted at the authors' discretion after manual review. By requesting access you agree to the model License (see the LICENSE file). In summary: use the model lawfully and responsibly; do not use it to identify, surveil, or de-anonymize individuals or to infer or extract sensitive personal information; do not redistribute or re-upload the weights; commercial or production use requires prior written permission; and attribute KSTU ASR (kstunlp) in any derived work.
Log in or Sign Up to review the conditions and access this model content.
kstu asr
Full fine-tune of facebook/mms-1b-all on Kyrgyz speech (real: kstunlp + Common Voice 21 + FLEURS-ky
train; plus synthetic augmentation). Greedy CTC decoding target_lang="kir".
Results — greedy, identical eval & text normalization
Two test sets: FLEURS-ky test (977 utts, read speech) and a held-out conversational test (1000 utts, real speech the baselines never saw).
| Model | Params | FLEURS WER | FLEURS CER | Conversational WER |
|---|---|---|---|---|
| Ours — MMS-1B full-FT | 1B | 12.93 (95% CI 12.2–13.7) | 3.08 | 16.17 |
MMS-1b-all zero-shot (Meta kir) |
1B | 17.13 | — | 27.02 |
| nineninesix/kyrgyz-whisper-medium | 769M | 18.06 | 5.27 | 43.45 |
| UlutSoftLLC/whisper-small-kyrgyz | 242M | 18.18 | 5.66 | 39.88 |
| kyrgyz-ai/AkylAI-STT-small * | 242M | 18.18 | 5.66 | 39.88 |
| nineninesix/kyrgyz-whisper-small | 242M | 18.98 | 5.07 | 38.68 |
| arfik/wav2vec2-large-mms-1b-kyrgyz | 1B | 19.60 | 5.17 | — |
#1 among all public Kyrgyz ASR — by ~5 WER points on FLEURS and ~22 points on real conversational speech The Whisper-based baselines collapse on conversational audio (~39–43%); the full fine-tune generalizes.
Usage
from transformers import Wav2Vec2ForCTC, AutoProcessor
import torch, soundfile as sf
proc = AutoProcessor.from_pretrained("kstunlp/kstu-asr"); proc.tokenizer.set_target_lang("kir")
model = Wav2Vec2ForCTC.from_pretrained("kstunlp/kstu-asr").eval()
wav, sr = sf.read("audio.wav") # 16 kHz mono
iv = proc(wav, sampling_rate=16000, return_tensors="pt").input_values
print(proc.batch_decode(torch.argmax(model(iv).logits, -1))[0])
License & access
This model is gated and released under a custom KSTU ASR Model License (other; see the LICENSE file). Access is granted at the authors' discretion after manual review and is open to requesters inside and outside KSTU. Use is permitted for lawful research, academic, and evaluation purposes. Do not use the model to identify, surveil, or de-anonymize individuals or to infer or extract sensitive personal information; do not redistribute or re-upload the weights; commercial or production use requires prior written permission; attribution to KSTU ASR (kstunlp) is required.
- Downloads last month
- 5