nur-dev/kk-ru-asr-tdt-accuracy

KK/RU ASR TDT Accuracy

Accuracy-focused Kazakh and Russian ASR/STT model packaged as a NeMo .nemo checkpoint.

This model is intended for high-quality offline or server-side transcription where accuracy is prioritized over minimum latency.

Files

• kk_ru_asr_tdt_accuracy.nemo - NeMo ASR checkpoint.
• examples/offline_asr.py - batch/file transcription example.
• examples/serve_fastapi.py - minimal HTTP service example.
• metadata/eval.json - evaluation summary.

Model Characteristics

• Task: automatic speech recognition / speech-to-text.
• Languages: Kazakh (kk) and Russian (ru).
• Format: NVIDIA NeMo .nemo.
• Decoder: RNNT/TDT.
• Recommended decoding for best accuracy: RNNT beam search with beam size 4.
• Fast decoding option: RNNT greedy or beam size 2.
• Input audio: mono 16 kHz PCM WAV or another format supported by your NeMo audio stack.

This is not a TTS model, not a diarization model, and not a cache-aware streaming checkpoint.

Evaluation

WER/CER on held-out local evaluation sets:

Decode	Set	Overall WER	Overall CER	KK WER	RU WER	RTF
Greedy	clean	3.321%	0.680%	2.351%	5.719%	0.0019
Greedy	original/noisy	10.549%	6.481%	13.221%	5.719%	0.0020
Beam 2	clean	3.257%	0.656%	2.308%	5.602%	0.0047
Beam 2	original/noisy	10.468%	6.454%	13.156%	5.607%	0.0048
Beam 4	clean	3.197%	0.644%	2.229%	5.588%	0.0121
Beam 4	original/noisy	10.430%	6.443%	13.107%	5.588%	0.0121

Recommended deployment default:

• Use beam_size=4 when accuracy is the priority.
• Use beam_size=2 when latency/cost matters.
• Use greedy only for maximum throughput.

Install

Use an environment with NVIDIA NeMo ASR support and CUDA-enabled PyTorch.

pip install "nemo_toolkit[asr]" huggingface_hub soundfile fastapi uvicorn python-multipart

Download

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="nur-dev/kk-ru-asr-tdt-accuracy",
    filename="kk_ru_asr_tdt_accuracy.nemo",
    token=True,
)
print(model_path)

Offline Inference

python examples/offline_asr.py \
  --model kk_ru_asr_tdt_accuracy.nemo \
  --audio /path/to/audio.wav \
  --strategy beam \
  --beam-size 4

Python:

from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel
from omegaconf import OmegaConf

model = EncDecHybridRNNTCTCBPEModel.restore_from("kk_ru_asr_tdt_accuracy.nemo")
model.freeze()

decoding_cfg = OmegaConf.create(OmegaConf.to_container(model.cfg.decoding, resolve=True))
decoding_cfg.strategy = "beam"
decoding_cfg.beam.beam_size = 4
decoding_cfg.beam.return_best_hypothesis = True
model.change_decoding_strategy(decoding_cfg=decoding_cfg, decoder_type="rnnt")

text = model.transcribe(["/path/to/audio.wav"], batch_size=1, return_hypotheses=False)[0]
print(text)

Serve With FastAPI

MODEL_PATH=kk_ru_asr_tdt_accuracy.nemo \
ASR_STRATEGY=beam \
ASR_BEAM_SIZE=4 \
uvicorn examples.serve_fastapi:app --host 0.0.0.0 --port 8000

Request:

curl -X POST "http://localhost:8000/transcribe" \
  -F "file=@/path/to/audio.wav"

Response:

{"text": "..."}

Best Practices

• Resample incoming audio to mono 16 kHz before inference.
• Use beam size 4 for final transcription.
• Use beam size 2 or greedy for interactive previews.
• Batch multiple files for throughput-oriented jobs.
• Do not use this checkpoint as a diarizer; combine it with a separate diarization model for speaker labels.
• For audio with many speakers or heavy overlap, use diarization and segment-level transcription before merging the final transcript.

Evaluation (independently measured)

Held-out public test sets, measured directly — not self-reported (seed 42, uniform multilingual-Whisper normalization). FLEURS test = 500 utterances/language; ISSAI KSC2 test = 1000 utterances (in-domain Kazakh, spanning crowd/parliament/podcasts/radio/talkshow).

Test set	Lang	WER (%)	CER (%)
FLEURS kk_kz	kk	11.48	6.04
FLEURS ru_ru	ru	12.75	6.99
FLEURS en_us	en	100.47	90.47
ISSAI KSC2	kk	9.75	3.50

Macro WER (kk/ru/en): 41.57% (unweighted mean; penalises models that do not cover all three languages).

Note. The card reports author clean-set KK 2.35 / RU 5.72. On held-out data: FLEURS kk 11.48 and KSC2 9.75 — still the best Kazakh in this account and robust across broadcast domains. It has no English: English audio is transcribed as Cyrillic phonetics, so EN WER exceeds 100%.

License & commercial use

Non-commercial use only (CC BY-NC 4.0). For commercial licensing or other inquiries, please reach out to the author, Nurgali Kadyrbek, on LinkedIn: https://www.linkedin.com/in/nurgali-kadyrbek-504260231/