kazakh-whisper-large-v3-turbo-robust-2epoch API & Inference Endpoint

Key Features

Robustness-tuned Kazakh ASR model
Based on Kazakh Whisper Large-v3 Turbo
Built on Whisper Large-v3 Turbo architecture
Fine-tuned for 2 robustness epochs
Trained with noise, reverb, speed, and volume perturbations
Ready-to-use Transformers checkpoint
Compatible with Hugging Face pipelines
Suitable for noisy Kazakh ASR experiments and evaluation
Useful for transcription, subtitles, voice assistants, and research

Important Note

This model is not intended to fully replace the original Kazakh Whisper Large-v3 Turbo in every situation.

Compared to the original model:

Clean FLEURS performance is slightly worse.
Robust/noisy evaluation performance is slightly better.
The main improvements appear under stronger additive noise conditions.

For clean high-quality audio, the original model remains a very strong choice.

For noisy or perturbed audio, this robust version may be preferable.

Benchmark Summary

Clean FLEURS Kazakh Test

Table with columns: Model, WER ↓, CER ↓
Model	WER ↓	CER ↓
Kazakh Whisper Large-v3 Turbo	11.74%	4.97%
Kazakh Whisper Large-v3 Turbo Robust 2epoch	11.90%	5.00%
Whisper Large-v3 Turbo	21.16%	6.12%
Wav2Vec2 XLSR Kazakh	21.76%	6.27%
Whisper Large-v3	33.10%	7.77%
Whisper Medium

Clean FLEURS performance is almost preserved, with a small WER increase compared to the original Kazakh Whisper Large-v3 Turbo model.

Robust FLEURS Kazakh

Robust FLEURS Kazakh was created from the Google FLEURS Kazakh test set by applying controlled perturbations to the original test audio.

Conditions:

clean
noise_10db
noise_5db
reverb
speed_0.9
speed_1.1
volume_low
volume_high

Robust FLEURS Results

Table with columns: Model, Clean WER ↓, Robust-only WER ↓, Robust-only CER ↓
Model	Clean WER ↓	Robust-only WER ↓	Robust-only CER ↓
Kazakh Whisper Large-v3 Turbo	11.76%	17.43%	6.84%
Kazakh Whisper Large-v3 Turbo Robust 2epoch	11.94%	17.33%	6.74%

The robust 2epoch model gives a small improvement on the robust-only subset:

Table with columns: Metric, Original, Robust 2epoch, Difference
Metric	Original	Robust 2epoch	Difference
Robust-only WER ↓	17.43%	17.33%	-0.10
Robust-only CER ↓	6.84%	6.74%	-0.10

The improvement is small and mainly appears under additive noise conditions.

Noise-only FLEURS Kazakh

A separate noise-only benchmark was created from Google FLEURS Kazakh test audio.

This benchmark focuses on synthetic stationary additive noise.

Conditions:

clean
white_20db
white_15db
white_10db
white_5db
white_0db
pink_20db
pink_15db
pink_10db
pink_5db
pink_0db

White noise is additive white Gaussian noise.

Pink noise is synthetic pink noise generated using approximate 1/sqrt(f) spectral shaping.

Noise-only FLEURS Results

Table with columns: Model, Clean WER ↓, Robust-only WER ↓, Robust-only CER ↓
Model	Clean WER ↓	Robust-only WER ↓	Robust-only CER ↓
Kazakh Whisper Large-v3 Turbo	11.76%	25.43%	9.88%
Kazakh Whisper Large-v3 Turbo Robust 2epoch	11.94%	25.15%	9.69%
Whisper Large-v3	33.18%	54.09%	18.54%
Whisper Medium	53.13%	73.87%

On the noise-only benchmark, the robust 2epoch model improves over the original model:

Table with columns: Metric, Original, Robust 2epoch, Difference
Metric	Original	Robust 2epoch	Difference
Robust-only WER ↓	25.43%	25.15%	-0.28
Robust-only CER ↓	9.88%	9.69%	-0.19

Detailed Noise-only WER

Table with columns: Condition, Kazakh Whisper Large-v3 Turbo, Robust 2epoch
Condition	Kazakh Whisper Large-v3 Turbo	Robust 2epoch
clean	11.76%	11.94%
white_20db	14.37%	14.51%
white_15db	17.67%	17.38%
white_10db	23.95%	23.92%
white_5db	35.33%	34.68%
white_0db	51.23%

Usage

Installation

bash

pip install transformers accelerate torch torchaudio soundfile

Quick Start

The model can be used directly with the Hugging Face pipeline() API:

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch") result = asr(    "audio.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

GPU Inference

For faster inference, use GPU inference with FP16:

python

import torchfrom transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch",    torch_dtype=torch.float16,    device="cuda") result = asr(    "audio.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

CPU Inference

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch",    device="cpu") result = asr(    "audio.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

Long Audio

For long audio files such as meetings, podcasts, interviews, and lectures, use chunked inference:

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch",    chunk_length_s=30,    batch_size=8) result = asr(    "meeting.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

Batch Processing

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch") files = [    "audio1.wav",    "audio2.wav",    "audio3.wav"] for file in files:    result = asr(        file,        generate_kwargs={            "language": "kk",            "task": "transcribe"        }    )     print(file, result["text"])

Direct Transformers Usage

The model can also be used directly with WhisperProcessor and WhisperForConditionalGeneration:

python

import torchimport torchaudiofrom transformers import WhisperProcessor, WhisperForConditionalGeneration model_id = "shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch" processor = WhisperProcessor.from_pretrained(model_id) model = WhisperForConditionalGeneration.from_pretrained(    model_id,    torch_dtype=torch.float16,    device_map="auto") audio_path = "audio.wav" speech_array, sampling_rate = torchaudio.load(audio_path) if sampling_rate != 16000:    resampler = torchaudio.transforms.Resample(sampling_rate, 16000)    speech_array = resampler(speech_array) audio_array = speech_array.squeeze().numpy() inputs = processor(    audio_array,    sampling_rate=16000,    return_tensors="pt") input_features = inputs.input_features.to(    model.device,    dtype=torch.float16) with torch.no_grad():    predicted_ids = model.generate(        input_features,        language="kk",        task="transcribe",        max_new_tokens=225    ) text = processor.batch_decode(    predicted_ids,    skip_special_tokens=True)[0] print("Prediction:", text)

Dataset Annotation

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch") audio_files = [    "sample1.wav",    "sample2.wav",    "sample3.wav"] annotations = [] for audio_file in audio_files:    result = asr(        audio_file,        generate_kwargs={            "language": "kk",            "task": "transcribe"        }    )     annotations.append({        "audio": audio_file,        "transcript": result["text"]    }) print(annotations)

Best Practices

For best performance:

Use 16 kHz audio when possible.
Use language="kk" and task="transcribe".
Segment very long recordings before transcription.
Use chunked inference for meetings, podcasts, interviews, and lectures.
Use GPU inference for large-scale workloads.
Evaluate on your own target-domain audio before production deployment.

Training Data

This model was fine-tuned from shyngys879/kazakh-whisper-large-v3-turbo using robustness-augmented Kazakh speech data.

The robustness training data was based on real-speech Kazakh data selected from issai/Kazakh_Speech_Corpus_2, excluding TTS audio.

The selected clean subset contained approximately 300 hours of real Kazakh speech.

It was augmented with multiple perturbation types, producing approximately 2,400 hours of robustness-focused training audio.

Augmentation conditions included:

clean
additive noise
reverberation
speed changes
volume changes

Intended Applications

This model can be used for:

Speech transcription
Subtitle generation
Podcast transcription
Interview transcription
Meeting transcription
Voice assistants
Noisy Kazakh ASR experiments
Dataset annotation
Educational applications
Call-center analytics
Kazakh NLP pipelines

Model Details

Table with columns: Field, Value
Field	Value
Model Name	Kazakh Whisper Large-v3 Turbo Robust 2epoch
Architecture	Whisper
Base Model	shyngys879/kazakh-whisper-large-v3-turbo
Original Base Architecture	Whisper Large-v3 Turbo
Language	Kazakh
Parameters	0.8B
Fine-tuning Type	Full model fine-tuning
Robustness Training Data	~300h clean real speech augmented to ~2,400h
Effective Robustness Epochs

Limitations

Performance may degrade on:

Heavy real-world background noise
Overlapping speakers
Strong accents or dialects
Code-switched Kazakh/Russian speech
Music-heavy recordings
Far-field audio
Call-center audio with compression artifacts
Very long audio without segmentation

Important evaluation limitation:

The noise-only benchmark uses synthetic white and pink noise. It is useful for controlled comparison, but it is not a full replacement for real-world noisy audio evaluation with babble noise, street noise, music, codec artifacts, and far-field microphones.

Citation

bibtex

@misc{sovetkhan2026kazakhwhisperrobust,  title={Kazakh Whisper Large-v3 Turbo Robust 2epoch},  author={Shyngys Sovetkhan},  year={2026},  howpublished={Hugging Face Model Hub},  url={https://huggingface.co/shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch}}

Original Kazakh Whisper Large-v3 Turbo:

text

shyngys879/kazakh-whisper-large-v3-turbo

Key Features

Robustness-tuned Kazakh ASR model
Based on Kazakh Whisper Large-v3 Turbo
Built on Whisper Large-v3 Turbo architecture
Fine-tuned for 2 robustness epochs
Trained with noise, reverb, speed, and volume perturbations
Ready-to-use Transformers checkpoint
Compatible with Hugging Face pipelines
Suitable for noisy Kazakh ASR experiments and evaluation
Useful for transcription, subtitles, voice assistants, and research

Important Note

This model is not intended to fully replace the original Kazakh Whisper Large-v3 Turbo in every situation.

Compared to the original model:

Clean FLEURS performance is slightly worse.
Robust/noisy evaluation performance is slightly better.
The main improvements appear under stronger additive noise conditions.

For clean high-quality audio, the original model remains a very strong choice.

For noisy or perturbed audio, this robust version may be preferable.

Benchmark Summary

Clean FLEURS Kazakh Test

Table with columns: Model, WER ↓, CER ↓
Model	WER ↓	CER ↓
Kazakh Whisper Large-v3 Turbo	11.74%	4.97%
Kazakh Whisper Large-v3 Turbo Robust 2epoch	11.90%	5.00%
Whisper Large-v3 Turbo	21.16%	6.12%
Wav2Vec2 XLSR Kazakh	21.76%	6.27%
Whisper Large-v3	33.10%	7.77%
Whisper Medium

Clean FLEURS performance is almost preserved, with a small WER increase compared to the original Kazakh Whisper Large-v3 Turbo model.

Robust FLEURS Kazakh

Robust FLEURS Kazakh was created from the Google FLEURS Kazakh test set by applying controlled perturbations to the original test audio.

Conditions:

clean
noise_10db
noise_5db
reverb
speed_0.9
speed_1.1
volume_low
volume_high

Robust FLEURS Results

Table with columns: Model, Clean WER ↓, Robust-only WER ↓, Robust-only CER ↓
Model	Clean WER ↓	Robust-only WER ↓	Robust-only CER ↓
Kazakh Whisper Large-v3 Turbo	11.76%	17.43%	6.84%
Kazakh Whisper Large-v3 Turbo Robust 2epoch	11.94%	17.33%	6.74%

The robust 2epoch model gives a small improvement on the robust-only subset:

Table with columns: Metric, Original, Robust 2epoch, Difference
Metric	Original	Robust 2epoch	Difference
Robust-only WER ↓	17.43%	17.33%	-0.10
Robust-only CER ↓	6.84%	6.74%	-0.10

The improvement is small and mainly appears under additive noise conditions.

Noise-only FLEURS Kazakh

A separate noise-only benchmark was created from Google FLEURS Kazakh test audio.

This benchmark focuses on synthetic stationary additive noise.

Conditions:

clean
white_20db
white_15db
white_10db
white_5db
white_0db
pink_20db
pink_15db
pink_10db
pink_5db
pink_0db

White noise is additive white Gaussian noise.

Pink noise is synthetic pink noise generated using approximate 1/sqrt(f) spectral shaping.

Noise-only FLEURS Results

Table with columns: Model, Clean WER ↓, Robust-only WER ↓, Robust-only CER ↓
Model	Clean WER ↓	Robust-only WER ↓	Robust-only CER ↓
Kazakh Whisper Large-v3 Turbo	11.76%	25.43%	9.88%
Kazakh Whisper Large-v3 Turbo Robust 2epoch	11.94%	25.15%	9.69%
Whisper Large-v3	33.18%	54.09%	18.54%
Whisper Medium	53.13%	73.87%

On the noise-only benchmark, the robust 2epoch model improves over the original model:

Table with columns: Metric, Original, Robust 2epoch, Difference
Metric	Original	Robust 2epoch	Difference
Robust-only WER ↓	25.43%	25.15%	-0.28
Robust-only CER ↓	9.88%	9.69%	-0.19

Detailed Noise-only WER

Table with columns: Condition, Kazakh Whisper Large-v3 Turbo, Robust 2epoch
Condition	Kazakh Whisper Large-v3 Turbo	Robust 2epoch
clean	11.76%	11.94%
white_20db	14.37%	14.51%
white_15db	17.67%	17.38%
white_10db	23.95%	23.92%
white_5db	35.33%	34.68%
white_0db	51.23%

Usage

Installation

bash

pip install transformers accelerate torch torchaudio soundfile

Quick Start

The model can be used directly with the Hugging Face pipeline() API:

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch") result = asr(    "audio.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

GPU Inference

For faster inference, use GPU inference with FP16:

python

import torchfrom transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch",    torch_dtype=torch.float16,    device="cuda") result = asr(    "audio.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

CPU Inference

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch",    device="cpu") result = asr(    "audio.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

Long Audio

For long audio files such as meetings, podcasts, interviews, and lectures, use chunked inference:

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch",    chunk_length_s=30,    batch_size=8) result = asr(    "meeting.wav",    generate_kwargs={        "language": "kk",        "task": "transcribe"    }) print(result["text"])

Batch Processing

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch") files = [    "audio1.wav",    "audio2.wav",    "audio3.wav"] for file in files:    result = asr(        file,        generate_kwargs={            "language": "kk",            "task": "transcribe"        }    )     print(file, result["text"])

Direct Transformers Usage

The model can also be used directly with WhisperProcessor and WhisperForConditionalGeneration:

python

import torchimport torchaudiofrom transformers import WhisperProcessor, WhisperForConditionalGeneration model_id = "shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch" processor = WhisperProcessor.from_pretrained(model_id) model = WhisperForConditionalGeneration.from_pretrained(    model_id,    torch_dtype=torch.float16,    device_map="auto") audio_path = "audio.wav" speech_array, sampling_rate = torchaudio.load(audio_path) if sampling_rate != 16000:    resampler = torchaudio.transforms.Resample(sampling_rate, 16000)    speech_array = resampler(speech_array) audio_array = speech_array.squeeze().numpy() inputs = processor(    audio_array,    sampling_rate=16000,    return_tensors="pt") input_features = inputs.input_features.to(    model.device,    dtype=torch.float16) with torch.no_grad():    predicted_ids = model.generate(        input_features,        language="kk",        task="transcribe",        max_new_tokens=225    ) text = processor.batch_decode(    predicted_ids,    skip_special_tokens=True)[0] print("Prediction:", text)

Dataset Annotation

python

from transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch") audio_files = [    "sample1.wav",    "sample2.wav",    "sample3.wav"] annotations = [] for audio_file in audio_files:    result = asr(        audio_file,        generate_kwargs={            "language": "kk",            "task": "transcribe"        }    )     annotations.append({        "audio": audio_file,        "transcript": result["text"]    }) print(annotations)

Best Practices

For best performance:

Use 16 kHz audio when possible.
Use language="kk" and task="transcribe".
Segment very long recordings before transcription.
Use chunked inference for meetings, podcasts, interviews, and lectures.
Use GPU inference for large-scale workloads.
Evaluate on your own target-domain audio before production deployment.

Training Data

This model was fine-tuned from shyngys879/kazakh-whisper-large-v3-turbo using robustness-augmented Kazakh speech data.

The robustness training data was based on real-speech Kazakh data selected from issai/Kazakh_Speech_Corpus_2, excluding TTS audio.

The selected clean subset contained approximately 300 hours of real Kazakh speech.

It was augmented with multiple perturbation types, producing approximately 2,400 hours of robustness-focused training audio.

Augmentation conditions included:

clean
additive noise
reverberation
speed changes
volume changes

Intended Applications

This model can be used for:

Speech transcription
Subtitle generation
Podcast transcription
Interview transcription
Meeting transcription
Voice assistants
Noisy Kazakh ASR experiments
Dataset annotation
Educational applications
Call-center analytics
Kazakh NLP pipelines

Model Details

Table with columns: Field, Value
Field	Value
Model Name	Kazakh Whisper Large-v3 Turbo Robust 2epoch
Architecture	Whisper
Base Model	shyngys879/kazakh-whisper-large-v3-turbo
Original Base Architecture	Whisper Large-v3 Turbo
Language	Kazakh
Parameters	0.8B
Fine-tuning Type	Full model fine-tuning
Robustness Training Data	~300h clean real speech augmented to ~2,400h
Effective Robustness Epochs

Limitations

Performance may degrade on:

Heavy real-world background noise
Overlapping speakers
Strong accents or dialects
Code-switched Kazakh/Russian speech
Music-heavy recordings
Far-field audio
Call-center audio with compression artifacts
Very long audio without segmentation

Important evaluation limitation:

Citation

bibtex

@misc{sovetkhan2026kazakhwhisperrobust,  title={Kazakh Whisper Large-v3 Turbo Robust 2epoch},  author={Shyngys Sovetkhan},  year={2026},  howpublished={Hugging Face Model Hub},  url={https://huggingface.co/shyngys879/kazakh-whisper-large-v3-turbo-robust-2epoch}}

Original Kazakh Whisper Large-v3 Turbo:

text

shyngys879/kazakh-whisper-large-v3-turbo

kazakh-whisper-large-v3-turbo-robust-2epoch

README

Key Features

Important Note

Benchmark Summary

Clean FLEURS Kazakh Test

Robust FLEURS Kazakh

Robust FLEURS Results

Noise-only FLEURS Kazakh

Noise-only FLEURS Results

Detailed Noise-only WER

Usage

Installation

Quick Start

GPU Inference

CPU Inference

Long Audio

Batch Processing

Direct Transformers Usage

Dataset Annotation

Best Practices

Training Data

Intended Applications

Model Details

Limitations

Citation

Related Model

Explore FriendliAI today

README

Key Features

Important Note

Benchmark Summary

Clean FLEURS Kazakh Test

Robust FLEURS Kazakh

Robust FLEURS Results

Noise-only FLEURS Kazakh

Noise-only FLEURS Results

Detailed Noise-only WER

Usage

Installation

Quick Start

GPU Inference

CPU Inference

Long Audio

Batch Processing

Direct Transformers Usage

Dataset Annotation

Best Practices

Training Data

Intended Applications

Model Details

Limitations

Citation

Related Model