🚨🚨🚨 ATTENTION! 🚨🚨🚨

Use an updated model: https://huggingface.co/Yehor/w2v-bert-uk-v2.1

---

w2v-bert-uk v1

Community

• Discord: https://bit.ly/discord-uds
• Speech Recognition: https://t.me/speech_recognition_uk
• Speech Synthesis: https://t.me/speech_synthesis_uk

See other Ukrainian models: https://github.com/egorsmkv/speech-recognition-uk

Google Colab

You can run this model using a Google Colab notebook: https://colab.research.google.com/drive/1QoKw2DWo5a5XYw870cfGE3dJf1WjZgrj?usp=sharing

Metrics

• AM (F16):
  • WER: 0.066 metric, 6.6%
  • CER: 0.013 metric, 1.34%
  • Accuracy on words: 93.4%
  • Accuracy on chars: 98.7%

Hyperparameters

This model was trained with the following hparams using 2 RTX A4000:

torchrun --standalone --nnodes=1 --nproc-per-node=2 ../train_w2v2_bert.py \
  --custom_set ~/cv10/train.csv \
  --custom_set_eval ~/cv10/test.csv \
  --num_train_epochs 15 \
  --tokenize_config . \
  --w2v2_bert_model facebook/w2v-bert-2.0 \
  --batch 4 \
  --num_proc 5 \
  --grad_accum 1 \
  --learning_rate 3e-5 \
  --logging_steps 20 \
  --eval_step 500 \
  --group_by_length \
  --attention_dropout 0.0 \
  --activation_dropout 0.05 \
  --feat_proj_dropout 0.05 \
  --feat_quantizer_dropout 0.0 \
  --hidden_dropout 0.05 \
  --layerdrop 0.0 \
  --final_dropout 0.0 \
  --mask_time_prob 0.0 \
  --mask_time_length 10 \
  --mask_feature_prob 0.0 \
  --mask_feature_length 10

Usage

# pip install -U torch soundfile transformers

import torch
import soundfile as sf
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor

# Config
model_name = 'Yehor/w2v-bert-2.0-uk'
device = 'cuda:1' # or cpu
sampling_rate = 16_000

# Load the model
asr_model = AutoModelForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2BertProcessor.from_pretrained(model_name)

paths = [
  'sample1.wav',
]

# Extract audio
audio_inputs = []
for path in paths:
  audio_input, _ = sf.read(path)
  audio_inputs.append(audio_input)

# Transcribe the audio
inputs = processor(audio_inputs, sampling_rate=sampling_rate).input_features
features = torch.tensor(inputs).to(device)

with torch.no_grad():
  logits = asr_model(features).logits

predicted_ids = torch.argmax(logits, dim=-1)
predictions = processor.batch_decode(predicted_ids)

# Log results
print('Predictions:')
print(predictions)

Cite this work

@misc {smoliakov_2025,
    author       = { {Smoliakov} },
    title        = { w2v-bert-uk (Revision e5a17ab) },
    year         = 2025,
    url          = { https://huggingface.co/Yehor/w2v-bert-uk },
    doi          = { 10.57967/hf/4560 },
    publisher    = { Hugging Face }
}