Introduction

AI4Kazakh just opened two Kazakh speech-recognition checkpoints on Hugging Face: ai4kazakh/qwen3-asr-0.6b-kazakh, created on 8 November 2026 at 16:20 UTC, and its larger sibling ai4kazakh/qwen3-asr-1.7b-kazakh, created a day earlier at 17:28 UTC on 7 November. Both are fine-tunes of Qwen's Qwen3-ASR models, both ship a full model.safetensors file, and both load today with Qwen3ASRModel.from_pretrained from the qwen-asr package. The licence is CC BY-NC 4.0 — free to use, not free to sell.

The official Qwen3-ASR-0.6B card lists 30 languages and 22 Chinese dialects. Kazakh is not among them. These drops are the practical unlock, the same way last week's Lao fine-tune opened a language the base card never advertised.

What Shipped

  • Vendor / model: AI4Kazakh — Qwen3-ASR-0.6B Kazakh and Qwen3-ASR-1.7B Kazakh
  • Modality: open-weight automatic speech recognition for Kazakh (kk)
  • Open vs closed: open weights, Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0)
  • Runnable today: qwen-asr / Qwen3ASRModel.from_pretrained on CUDA in bfloat16; ungated Hugging Face downloads
  • Concrete fact: both were trained on about 960k utterances. On the ISSAI KSC2 test set (9,351 clips) the 1.7B scores 11.04% WER and 3.72% CER overall; the 0.6B scores 12.92% WER and 4.27% CER

AI4Kazakh describes itself as a community building open models, datasets, and tools for AI in Kazakhstan. These two cards are the org's current model listing. Hugging Face reports about 0.8B BF16 parameters on the smaller checkpoint and about 2B on the larger one. Each repo includes tokenizer files plus a single model.safetensors — a full fine-tune, not a LoRA adapter sitting on top of the base weights.

The training mix is the same for both sizes:

  • ISSAI Kazakh Speech Corpus 2: 627,824 utterances
  • Alfarabi STT (Farabi Lab): 173,610 utterances
  • NCSpeech YO-CPT-kk: 156,903 utterances

Transcripts were normalized to lowercase Kazakh Cyrillic letters and single spaces only — no punctuation, no digits. Training used the base model's multilingual prompt format with a language Kazakh<asr_text> prefix.

How to Run It

The 0.6B card's own snippet (swap the repo id for the 1.7B checkpoint):

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "ai4kazakh/qwen3-asr-0.6b-kazakh",
    dtype=torch.bfloat16,
    device_map="cuda:0",
)

results = model.transcribe(audio=["sample.wav"])
print(results[0].text)
# output: lowercase Kazakh text without punctuation or fillers

Batch inference takes a list of paths. The card recommends 16 kHz mono WAV:

results = model.transcribe(audio=["clip1.wav", "clip2.wav", "clip3.wav"])
texts = [r.text for r in results]

Install with pip install -U qwen-asr if you do not already have the official Qwen3-ASR toolkit.

The Numbers

WER and CER were computed with Hugging Face evaluate (wer, cer) after lowercasing and stripping fillers such as ааа, мхм, and мм from both the reference and the prediction.

On ISSAI KSC2 (9,351 clips), overall:

  • 1.7B: 11.04% WER, 3.72% CER
  • 0.6B: 12.92% WER, 4.27% CER

Read and broadcast speech is the easy end. TV news is the cleanest source (1.7B 6.73% WER / 1.72% CER; 0.6B 7.74% / 2.01% on 2,618 clips). Crowdsourced clips follow (1.7B 7.40% / 1.84%; 0.6B 9.12% / 2.22% on 3,328 clips). Parliament sits close behind.

Spontaneous, multi-speaker, noisier audio is harder, and the cards say so plainly:

  • Podcasts (1,547 clips): 1.7B 19.47% WER / 8.27% CER; 0.6B 22.64% / 9.36%
  • Radio (702 clips): 1.7B 17.37% / 6.68%; 0.6B 19.43% / 7.59%
  • Talk shows (383 clips): 1.7B 17.59% / 7.94%; 0.6B 18.97% / 8.29%

An external Alfarabi STT test set of 30,638 clips tells a slightly different story. There the smaller model wins on word error (9.67% WER / 2.56% CER) while the 1.7B lands at 10.86% WER / 2.42% CER. If you care more about letter accuracy on that set, the larger model still edges CER. If you care about words, the 0.6B is the better pick on Alfarabi — and it is the lighter download.

Limits

  • The weights are CC BY-NC 4.0, so do not put the model itself into a commercial product without a separate permission the cards do not grant
  • Output is always lowercase and has no punctuation. Restore casing and stops yourself if you need a readable transcript
  • The models were trained and evaluated on 16 kHz mono audio; resample anything else first
  • Podcasts and radio remain the weakest sources — roughly 17–23% WER depending on size
  • These are speech-to-text fine-tunes, not voice generators. They do not map to ArtRealmAI Gen

Original Source

Qwen3-ASR-0.6B Kazakh on Hugging Face

The larger sibling is Qwen3-ASR-1.7B Kazakh.

Conclusion

If you needed an open Kazakh path on the Qwen3-ASR stack, it is here: two ungated checkpoints, the same ~960k-utterance mix, published scores on KSC2 and Alfarabi, and a one-block qwen-asr load. Start with the 0.6B if you want the lighter file; move to the 1.7B when KSC2's overall word error matters more. Keep the CC BY-NC 4.0 limit in mind, send 16 kHz mono, and expect the transcript back in lowercase Kazakh without punctuation.

—Anabel ♡