nur-dev/tynda-stt-4L

Tynda STT 4L

Tynda (Тыңда — "Listen" in Kazakh) is a multilingual speech-to-text model supporting 4 languages of Central Asia and beyond.

Supported Languages

Language	Code
Kazakh	kk
Russian	ru
English	en
Uzbek	uz

Model Details

• Architecture: Whisper Large V3 (1.55B parameters)
• Task: Automatic Speech Recognition / Speech-to-Text
• Audio Input: 16kHz mono WAV
• Max Duration: 30 seconds per segment

Usage

import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "nur-dev/tynda-stt-4L"
device = "cuda:0" if torch.cuda.is_available() else "cpu"

model = WhisperForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.float16
).to(device)

# Choose language: "kazakh", "russian", "english", or "uzbek"
processor = WhisperProcessor.from_pretrained(
    "openai/whisper-large-v3", language="kazakh", task="transcribe"
)

# Load your audio (16kHz mono)
import soundfile as sf
audio, sr = sf.read("audio.wav", dtype="float32")

inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
features = inputs.input_features.to(device, dtype=torch.float16)

forced_ids = processor.get_decoder_prompt_ids(language="kazakh", task="transcribe")

with torch.no_grad():
    predicted_ids = model.generate(
        features,
        forced_decoder_ids=forced_ids,
        max_new_tokens=200,
    )

text = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(text)

Using with pipeline

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="nur-dev/tynda-stt-4L",
    torch_dtype="float16",
    device="cuda:0",
)

result = pipe(
    "audio.wav",
    generate_kwargs={"language": "kazakh", "task": "transcribe"},
)
print(result["text"])

License

This model is released under CC BY-NC 4.0. It is free for non-commercial use. For commercial licensing, please contact the authors.

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	24.80	10.81
FLEURS ru_ru	ru	11.49	6.46
FLEURS en_us	en	6.33	3.41
ISSAI KSC2	kk	30.60	12.54

Macro WER (kk/ru/en): 14.21% (unweighted mean; penalises models that do not cover all three languages).

Note. The card reports no numbers. Best Russian and English in this account; Kazakh is the weak spot (FLEURS 24.8, and KSC2 30.6 on in-domain broadcast speech).

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/