Fine-tuned XLSR-53-russian large model for speech recognition in Macedonian
Authors:
Dejan Porjazovski
Ilina Jakimovska
Ordan Chukaliev
Nikola Stikov
This collaboration is part of the activities of the Center for Advanced Interdisciplinary Research (CAIR) at UKIM.
Note
This is an older version (Buki 1.0). It is recommended to use the latest version, trained with much more data: Macedonian-ASR/buki-wav2vec2-2.0
Data used for training
The model is trained on around 60 hours of Macedonian speech.
In training of the model, we used the following data sources:
Digital Archive for Ethnological and Anthropological Resources (DAEAR) at the Institutе of Ethnology and Anthropology, PMF, UKIM.
Audio version of the international journal "EthnoAnthropoZoom" at the Institutе of Ethnology and Anthropology, PMF, UKIM.
The podcast "Обични луѓе" by Ilina Jakimovska.
The scientific videos from the series "Наука за деца", foundation KANTAROT.
Macedonian version of the Mozilla Common Voice (version 18).
Model description
This model is an attention-based encoder-decoder (AED). The encoder is a Wav2vec2 model and the decoder is RNN-based.
Results
The results are reported on all the test sets combined and without an external language model.
WER: 13.77
CER: 5.03
Usage
The model is developed using the SpeechBrain toolkit. To use it, you need to install SpeechBrain with:
pip install speechbrain
SpeechBrain relies on the Transformers library, therefore you need install it:
pip install transformers
An external py_module_file=custom_interface.py is used as an external Predictor class into this HF repos. We use the foreign_class function from speechbrain.pretrained.interfaces that allows you to load your custom model.
train.py file contains the functions necessary for training the model and hyperparams.yaml contains the hyperparameters. For more details about training the model, refer to the SpeechBrain documentation.
