Slovak KKY FAV

wav2vec2-base-sk-17k

This is a monolingual Slovak Wav2Vec 2.0 base model pre-trained from 17 thousand of hours of Slovak speech.

This model does not have a tokenizer as it was pretrained on audio alone. In order to use this model for speech recognition, a tokenizer should be created, and the model should be fine-tuned on labeled data.

The model was initialized from Czech pre-trained model fav-kky/wav2vec2-base-cs-80k-ClTRUS. We found this cross-language transfer learning approach better than pre-training from scratch. See our paper for details.

Pretraining data

Almost 18 thousand hours of unlabeled Slovak speech:

Usage

Inputs must be 16kHz mono audio files.

This model can be used e.g. to extract per-frame contextual embeddings from audio:

from transformers import Wav2Vec2Model, Wav2Vec2FeatureExtractor
import torchaudio

feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("fav-kky/wav2vec2-base-sk-17k")
model = Wav2Vec2Model.from_pretrained("fav-kky/wav2vec2-base-sk-17k")

speech_array, sampling_rate = torchaudio.load("/path/to/audio/file.wav")
inputs = feature_extractor(
    speech_array, 
    sampling_rate=16_000, 
    return_tensors="pt"
)["input_values"][0]

output = model(inputs)
embeddings = output.last_hidden_state.detach().numpy()[0]

Speech recognition results

After fine-tuning, the model scored the following results on public datasets:

See our paper for details.

Paper

The preprint of our paper (accepted to TSD 2023) is available at https://arxiv.org/abs/2306.04399.

Citation

If you find this model useful, please cite our paper:

@inproceedings{wav2vec2-base-sk-17k,
  title = {{Transfer Learning of Transformer-based Speech Recognition Models from Czech to Slovak}},
  author = {
    Jan Lehe\v{c}ka and 
    Josef V. Psutka and 
    Josef Psutka
  },
  booktitle = {{Text, Speech, and Dialogue}},
  publisher = {{Springer International Publishing}},
  year = {2023},
  note = {(in press)},
  url = {https://arxiv.org/abs/2306.04399},
}

Related papers

Related models