Downloads · 30 days
17
9% of all-time downloads
KSE-RESEARCH-Group/whisper-large-v3-dido-yvanchyk-v2
whisper-large-v3-dido-yvanchyk-v2 is a automatic speech recognition model from KSE-RESEARCH-Group. Use it when you need speech turned into text. The card lists the license as mit.
This model is a fine-tuned version of openai/whisper-large-v3 on the Dido Yvanchyk Audio Dataset v2 for Ukrainian speech recognition.
Downloads · 30 days
17
9% of all-time downloads
All-time downloads
184
Public
Parameters
1.5B
6.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors6.2 GB · 100%
From the Hugging Face model README
This model is a fine-tuned version of openai/whisper-large-v3 on the Dido Yvanchyk Audio Dataset v2 for Ukrainian speech recognition.
Fine-tune openai/whisper-large-v3 on Dido-Yvanchyk dataset
| Property | Value |
|---|---|
| Dataset | KSE-RESEARCH-Group/Dido-Yvanchyk-Audio-Dataset-v2 |
| Training samples | 6811 |
| Validation samples | 757 |
| Test samples | 842 |
| Language | Ukrainian |
| Max token length | 448 |
| Parameter | Value |
|---|---|
| Base model | openai/whisper-large-v3 |
| Learning rate | 1e-05 |
| Warmup steps | 500 |
| Max steps | 8000 |
| Batch size (per device) | 4 |
| Gradient accumulation steps | 4 |
| Effective batch size | 16 |
| FP16 | True |
| Gradient checkpointing | False |
| Eval strategy | steps |
| Eval/Save steps | 500 |
| Metric for best model | cer |
The model was trained for 8000 steps with evaluation every 500 steps. The best checkpoint was selected based on the lowest CER.
| Step | Train Loss | Eval Loss | Eval CER (%) | Eval WER (%) |
|---|---|---|---|---|
| 500 | 0.15 | 0.211 | 5.36 | 16.24 |
| 1000 | 0.093 | 0.196 | 4.64 | 14.34 |
| 1500 | 0.06 | 0.205 | 4.19 | 14.43 |
| 2000 | 0.044 | 0.211 | 4.19 | 13.94 |
| 2500 | 0.021 | 0.227 | 4.09 | 13.04 |
| 3000 | 0.011 | 0.238 | 3.93 | 13.06 |
| 3500 | 0.005 | 0.25 | 3.97 | 13.19 |
| 4000 | 0.003 | 0.253 | 3.84 | 13.08 |
| 4500 | 0.002 | 0.255 | 3.85 | 12.74 |
| 5000 | 0.002 | 0.267 | 3.86 | 12.61 |
| 5500 | 0.001 | 0.275 | 3.89 | 12.82 |
| 6000 | 0.001 | 0.274 | 3.73 | 12.44 |
| 6500 | 0.001 | 0.28 | 3.87 | 12.62 |
| 7000 | 0.0 | 0.286 | 3.71 | 12.41 |
| 7500 | 0.0 | 0.29 | 3.73 | 12.42 |
| 8000 | 0.0 | 0.291 | 3.73 | 12.4 |
Best Model Checkpoint: Step 7000
| Metric | Value |
|---|---|
| CER | 3.73% |
| WER | 12.4% |
| Eval Loss | 0.291 |
| Metric | Value |
|---|---|
| CER | 3.9% |
| WER | 13.2% |
from transformers import pipeline
import torch
device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipe = pipeline(
"automatic-speech-recognition",
model="KSE-RESEARCH-Group/whisper-large-v3-dido-yvanchyk-v2",
device=device,
)
result = pipe(
"path/to/audio.wav",
generate_kwargs={
"task": "transcribe",
"language": "ukrainian",
},
chunk_length_s=30,
)
print(result["text"])
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch
model_id = "KSE-RESEARCH-Group/whisper-large-v3-dido-yvanchyk-v2"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)
# Move to GPU if available
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = model.to(device)
# Process audio (audio_array should be a numpy array at 16kHz)
input_features = processor(
audio_array,
sampling_rate=16000,
return_tensors="pt"
).input_features.to(device)
# Generate transcription
predicted_ids = model.generate(input_features)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)
| Component | Specification |
|---|---|
| GPU | NVIDIA GeForce RTX 4090 |
| GPU Memory | 47.38 GB |
| GPU Count | 1 |
| CUDA Compute Capability | 8.9 |
| Package | Version |
|---|---|
| Python | 3.12.12 |
| PyTorch | 2.8.0+cu128 |
| CUDA | 12.8 |
| Transformers | 4.52.4 |
| Datasets | 3.6.0 |
| Evaluate | 0.4.3 |
| Metric | Value |
|---|---|
| Total training time | 8:55:01.513110 |
| Training started | 2025-12-26 02:55:57 |
| Training completed | 2025-12-26 11:50:58 |
| Property | Value |
|---|---|
| Experiment ID | whisper-large-v3-011 |
| WandB Project | dido-yvanchik-stt |
| WandB Run | whisper-large-v3-011 |
If you use this model, please cite:
@misc{KSE-RESEARCH-Group-whisper-large-v3-dido-yvanchyk-v2,
author = {KSE-RESEARCH-Group},
title = {whisper-large-v3 - Fine-tuned for Ukrainian ASR},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/KSE-RESEARCH-Group/whisper-large-v3-dido-yvanchyk-v2}
}
This model is released under the MIT license.