Kazakh-Russian ASR Whisper Small LoRA

This repository contains a LoRA adapter for openai/whisper-small, fine-tuned for Kazakh and Kazakh-Russian mixed-speech automatic speech recognition.

The model is intended for speech-to-text experiments where Kazakh is the main language, while Russian words or short Russian phrases may appear naturally inside the utterance.

This repository does not contain a fully merged Whisper model.
It contains a PEFT/LoRA adapter that must be loaded together with the original openai/whisper-small base model.

Model Details

Field	Value
Base model	`openai/whisper-small`
Adaptation method	LoRA / PEFT
Model type	Whisper encoder-decoder ASR model
Task	Automatic Speech Recognition / Speech-to-Text
Languages	Kazakh, Russian, Kazakh-Russian mixed speech
Repository type	LoRA adapter
Project context	Academic thesis research

LoRA Configuration

Parameter	Value
Rank `r`	64
Alpha `α`	128
Dropout	0.05
Target modules	`q_proj`, `v_proj`, `k_proj`, `out_proj`, `fc1`, `fc2`

Recommended Use

This model is mainly intended for:

academic ASR research;
Kazakh speech recognition experiments;
Kazakh-Russian mixed-speech transcription;
code-switching ASR evaluation;
comparison with other KRASR ASR models;
reproducibility of thesis experiments;
demonstration in speech-to-text applications.

For best results, use it on speech where Kazakh is dominant and Russian appears as inserted words or phrases.

Quick Start

Install the required libraries:

pip install -U transformers peft accelerate torch librosa soundfile evaluate tqdm

Run inference on one audio file

import torch
import librosa
from peft import PeftModel
from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline

base_model_id = "openai/whisper-small"
adapter_id = "KRASR/kazakh-russian-asr-whisper-small-lora"
audio_path = "audio.wav"

device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipeline_device = 0 if torch.cuda.is_available() else -1
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

processor = WhisperProcessor.from_pretrained(base_model_id)

base_model = WhisperForConditionalGeneration.from_pretrained(
    base_model_id,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)

model = PeftModel.from_pretrained(base_model, adapter_id)
model = model.merge_and_unload()
model.to(device)
model.eval()

asr = pipeline(
    task="automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=pipeline_device,
)

audio, sr = librosa.load(audio_path, sr=16000, mono=True)

result = asr(
    {"array": audio, "sampling_rate": sr},
    generate_kwargs={
        "task": "transcribe",
        "language": "kazakh",
        "num_beams": 3,
        "no_repeat_ngram_size": 4,
        "repetition_penalty": 1.12,
    },
)

print(result["text"])

Decoding Notes

For normal testing, avoid using a fixed max_new_tokens value such as 96.
A fixed limit can accidentally cut off longer transcriptions.

A good starting point for Kazakh-dominant mixed speech is:

generate_kwargs = {
    "task": "transcribe",
    "language": "kazakh",
    "num_beams": 3,
    "no_repeat_ngram_size": 4,
    "repetition_penalty": 1.12,
}

Why forced Kazakh decoding?

In mixed Kazakh-Russian speech, automatic language detection can be unstable, especially on short utterances.
If the audio is mostly Kazakh with Russian insertions, forcing Kazakh decoding usually keeps the transcription closer to the target speech domain.

Russian words may still appear in the output when the model recognizes them from the audio.

Optional dynamic token limit for apps or batch evaluation

For applications or controlled batch evaluation, a dynamic limit is safer than one fixed value.
The following helper follows the same idea as the KRASR demo module: short clips receive a smaller output limit, while longer clips receive a larger limit.

def build_generate_kwargs(
    audio_duration_sec: float | None,
    language: str | None = "kazakh",
    num_beams: int = 3,
) -> dict:
    if audio_duration_sec is None:
        max_new_tokens = 96
    elif audio_duration_sec <= 5:
        max_new_tokens = 40
    elif audio_duration_sec <= 10:
        max_new_tokens = 64
    elif audio_duration_sec <= 15:
        max_new_tokens = 80
    elif audio_duration_sec <= 20:
        max_new_tokens = 96
    elif audio_duration_sec <= 25:
        max_new_tokens = 112
    else:
        max_new_tokens = 128

    generate_kwargs = {
        "task": "transcribe",
        "num_beams": int(num_beams),
        "max_new_tokens": max_new_tokens,
        "no_repeat_ngram_size": 4,
        "repetition_penalty": 1.12,
    }

    if language is not None:
        generate_kwargs["language"] = language

    return generate_kwargs

Use this only when you specifically need output-length control. For ordinary one-file testing, starting without max_new_tokens is usually simpler.

Evaluation Results

Evaluation set	WER	CER	Notes
Test-MIXED	0.5626	0.3308	Main Kazakh-Russian mixed-speech test set
Test-KK	0.5385	-	Internal pure Kazakh test set
Test-RU	0.7879	-	Internal pure Russian test set
FLEURS-KK	0.7603	-	External Kazakh benchmark
FLEURS-RU	0.7740	-	External Russian benchmark

The model significantly improved over the original Whisper-Small baseline on the main mixed-speech test set, but it is not the strongest model in the KRASR collection. Russian-only recognition became less stable after adaptation to Kazakh-dominant mixed speech.

The thesis evaluation used beam-search-based decoding and output-length control for the main fine-tuned Whisper-Small comparison.

Batch Evaluation Example

The following example shows how to test the model on a simple JSONL manifest.

Expected manifest format:

{"audio": "path/to/audio.wav", "text": "reference transcription"}

Evaluation script:

import json
import torch
import librosa
import evaluate
from tqdm import tqdm
from peft import PeftModel
from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline

base_model_id = "openai/whisper-small"
adapter_id = "KRASR/kazakh-russian-asr-whisper-small-lora"
manifest_path = "test_manifest.jsonl"

device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipeline_device = 0 if torch.cuda.is_available() else -1
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

wer_metric = evaluate.load("wer")
cer_metric = evaluate.load("cer")

def build_generate_kwargs(audio_duration_sec, language="kazakh", num_beams=3):
    if audio_duration_sec is None:
        max_new_tokens = 96
    elif audio_duration_sec <= 5:
        max_new_tokens = 40
    elif audio_duration_sec <= 10:
        max_new_tokens = 64
    elif audio_duration_sec <= 15:
        max_new_tokens = 80
    elif audio_duration_sec <= 20:
        max_new_tokens = 96
    elif audio_duration_sec <= 25:
        max_new_tokens = 112
    else:
        max_new_tokens = 128

    generate_kwargs = {
        "task": "transcribe",
        "num_beams": int(num_beams),
        "max_new_tokens": max_new_tokens,
        "no_repeat_ngram_size": 4,
        "repetition_penalty": 1.12,
    }

    if language is not None:
        generate_kwargs["language"] = language

    return generate_kwargs

processor = WhisperProcessor.from_pretrained(base_model_id)

base_model = WhisperForConditionalGeneration.from_pretrained(
    base_model_id,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)

model = PeftModel.from_pretrained(base_model, adapter_id)
model = model.merge_and_unload()
model.to(device)
model.eval()

asr = pipeline(
    task="automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=pipeline_device,
)

predictions = []
references = []

with open(manifest_path, "r", encoding="utf-8") as f:
    rows = [json.loads(line) for line in f]

for row in tqdm(rows):
    audio, sr = librosa.load(row["audio"], sr=16000, mono=True)
    duration_sec = len(audio) / sr

    result = asr(
        {"array": audio, "sampling_rate": sr},
        generate_kwargs=build_generate_kwargs(
            audio_duration_sec=duration_sec,
            language="kazakh",
            num_beams=3,
        ),
    )

    predictions.append(result["text"])
    references.append(row["text"])

wer = wer_metric.compute(predictions=predictions, references=references)
cer = cer_metric.compute(predictions=predictions, references=references)

print(f"WER: {wer:.4f}")
print(f"CER: {cer:.4f}")

Training Data

The model was fine-tuned using the KRASR/kazakh-russian-asr-dataset, prepared for Kazakh and Kazakh-Russian mixed-speech ASR experiments.

The dataset preparation workflow included:

source selection;
audio segmentation;
transcription review;
text normalization;
train/validation/test split preparation;
evaluation setup for mixed-language ASR.

The dataset was prepared for speech recognition only. It was not designed for speaker identification, biometric analysis, or demographic classification.

Preprocessing

Audio and text were prepared using a consistent ASR preprocessing pipeline.

Audio preprocessing:

mono audio;
16 kHz sampling rate;
short-segment ASR setting.

Text normalization included:

lowercasing;
whitespace normalization;
punctuation cleanup;
preservation of Kazakh-specific letters;
preservation of Russian words in mixed utterances;
removal of formatting noise that does not affect transcription meaning.

Known Limitations

The model may make errors on:

very short audio clips;
noisy recordings;
overlapping speech;
informal conversational speech;
rare names, places, and domain-specific terms;
silent or low-quality audio.

Like other Whisper-based models, it may sometimes produce extra words or hallucinated text, especially when the input audio is too short, unclear, or contains long silence.

Out-of-Scope Use

This model is not intended for:

speaker identification;
biometric profiling;
demographic classification;
surveillance or tracking of individuals;
high-stakes decision-making systems;
production deployment without additional validation.

Project Context

KRASR was created as part of an academic thesis project on automatic Kazakh speech-to-text conversion using fine-tuned multilingual ASR models.

The project compares Whisper-Small baseline, Whisper-Small LoRA, Whisper-Small full fine-tuning, XLS-R 1B CTC, and Whisper Large-v3 LoRA on Kazakh, Russian, and Kazakh-Russian mixed speech.

Related Repositories

KRASR/kazakh-russian-asr-dataset
KRASR/kazakh-russian-asr-whisper-small-lora
KRASR/kazakh-russian-asr-whisper-small-full-ft
KRASR/kazakh-russian-asr-whisper-large-v3-lora
KRASR/kazakh-russian-asr-xls-r-1b-ctc
KRASR/kazakh-russian-speech-to-text-module

Citation

There is no formal publication for this model yet.

If you use this model or dataset in academic work, please cite or mention the KRASR Hugging Face repository and the related thesis project:

@misc{krasr_whisper_small_lora_2026,
  title        = {Kazakh-Russian ASR Whisper Small LoRA},
  author       = {Mukhambet, Madiyar and Makhmud, Danial},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/KRASR/kazakh-russian-asr-whisper-small-lora}},
  note         = {LoRA adapter for Kazakh and Kazakh-Russian mixed-speech ASR}
}

Related thesis project:

Madiyar Mukhambet and Danial Makhmud.
Development of a Software Module for Automatic Kazakh Speech-to-Text Conversion Based on Fine-Tuned Whisper-Small Model.
Astana IT University, 2026.

Downloads last month: 4

Model tree for KRASR/kazakh-russian-asr-whisper-small-lora

Base model

openai/whisper-small

Adapter

(242)

this model