Model Details
Table with columns: Field, Value| Field | Value |
|---|
| Base model | openai/whisper-small |
| Adaptation method | Full fine-tuning |
| Model type | Whisper encoder-decoder ASR model |
| Task | Automatic Speech Recognition / Speech-to-Text |
| Languages | Kazakh, Russian, Kazakh-Russian mixed speech |
| Repository type | Full fine-tuned model |
| Project context | Academic thesis research |
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 Whisper-Small LoRA;
- reproducibility of thesis experiments;
- demonstration in speech-to-text applications.
This model is useful when you want to test a fully updated Whisper-Small checkpoint rather than a parameter-efficient LoRA adapter.
For stronger recognition quality, see the larger comparative models in the KRASR collection, especially KRASR/kazakh-russian-asr-whisper-large-v3-lora.
Quick Start
Install the required libraries:
pip install -U transformers accelerate torch librosa soundfile evaluate tqdm
Simple pipeline inference
import torch
from transformers import pipeline
model_id = "KRASR/kazakh-russian-asr-whisper-small-full-ft"
device = 0 if torch.cuda.is_available() else -1
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
asr = pipeline(
task="automatic-speech-recognition",
model=model_id,
torch_dtype=torch_dtype,
device=device,
)
result = asr(
"audio.wav",
generate_kwargs={
"task": "transcribe",
"language": "kazakh",
"num_beams": 3,
"no_repeat_ngram_size": 4,
"repetition_penalty": 1.12,
},
)
print(result["text"])
Manual Loading Example
import torch
import librosa
from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline
model_id = "KRASR/kazakh-russian-asr-whisper-small-full-ft"
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(model_id)
model = WhisperForConditionalGeneration.from_pretrained(
model_id,
torch_dtype=torch_dtype,
low_cpu_mem_usage=True,
)
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 output limit can be useful.
This follows the same idea as the KRASR demo module: short clips receive a smaller output limit, while longer clips receive a larger one.
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
Use this only when you specifically need output-length control.
For ordinary one-file testing, starting without max_new_tokens is usually simpler.
Training Configuration
Table with columns: Parameter, Value| Parameter | Value |
|---|
| Base model | openai/whisper-small |
| Training manifest | train_all.jsonl |
| Validation manifest | val_all.jsonl |
| Sampling rate | 16 kHz |
| Maximum audio duration | 30 seconds |
| Number of epochs | 8 |
| Learning rate | 1e-5 |
Full fine-tuning updates all Whisper-Small parameters. This gives the model more freedom to adapt to the target dataset, but it requires more compute than LoRA adaptation.
Evaluation Results
Main mixed-speech evaluation
Table with columns: Model, Decoding, WER, CER, MER, WIL, Hyp/ref, Possible hallucination-like cases| Model | Decoding | WER | CER | MER | WIL | Hyp/ref | Possible hallucination-like cases |
|---|
| Whisper-Small baseline | Greedy | 1.0852 | 0.8256 | 0.9254 | - | 0.8780 | - |
| Whisper-Small LoRA | Duration-based token limit + Beam3 | 0.5626 | 0.3308 |
The full fine-tuned model achieved slightly better WER and CER than Whisper-Small LoRA on the main Test-MIXED set, but the difference was small.
Effect of beam search on Whisper-Small Full FT
Table with columns: Decoding, WER, CER, Insertions, Hyp/ref, Possible hallucination-like cases| Decoding | WER | CER | Insertions | Hyp/ref | Possible hallucination-like cases |
|---|
| Greedy | 0.5828 | 0.3311 | 1493 | 0.9073 | 11 |
| Beam3 | 0.5581 | 0.3219 | 1204 | 0.8882 | 5 |
| Beam5 | 0.5551 |
Beam search improved the full fine-tuned model compared with greedy decoding. Beam5 gave the best observed Whisper-Small Full FT result on Test-MIXED, while Beam3 was used in the main comparison for consistency.
Pure-language and external benchmark behavior
Table with columns: Evaluation set, WER| Evaluation set | WER |
|---|
| Test-KK | 0.5393 |
| Test-RU | 0.7192 |
| FLEURS-KK | 0.7609 |
| FLEURS-RU | 0.7419 |
The model improved internal Kazakh recognition compared with the original Whisper-Small baseline, but Russian-only recognition became less stable after adaptation to Kazakh-dominant mixed speech.
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;
- long Russian segments inside Kazakh-dominant speech;
- 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.
The model is stronger than the original Whisper-Small baseline on the main mixed-speech evaluation set, but it remains weaker than the larger or alternative KRASR comparative models such as Whisper Large-v3 LoRA and XLS-R 1B CTC.
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.
KRASR/kazakh-russian-asr-datasetKRASR/kazakh-russian-asr-whisper-small-loraKRASR/kazakh-russian-asr-whisper-large-v3-loraKRASR/kazakh-russian-asr-xls-r-1b-ctcKRASR/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_full_ft_2026,
title = {Kazakh-Russian ASR Whisper Small Full Fine-Tuning},
author = {Mukhambet, Madiyar and Makhmud, Danial},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/KRASR/kazakh-russian-asr-whisper-small-full-ft}},
note = {Fully fine-tuned Whisper-Small model 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.