Automatic Speech Recognition for Belarusian language
Fine-tuned version of facebook/wav2vec2-base on mozilla-foundation/common_voice_8_0 be dataset.
Train, Dev, Test splits were used as they are present in the dataset. No additional data was used from Validated split,
only 1 voicing of each sentence was used - the way the data was split by CommonVoice CorporaCreator.
To build a better model one can use additional voicings from Validated split for sentences already present in Train, Dev, Test splits,
i.e. enlarge mentioned splits.
Language model was built using KenLM.
5-gram Language model was built on sentences from Train + (Other - Dev - Test) splits of mozilla-foundation/common_voice_8_0 be dataset.
Source code is available here.
Run model in a browser
This page contains interactive demo widget that lets you test this model right in a browser.
However, this widget uses Acoustic model only without Language model that significantly improves overall performance.
You can play with full pipeline of Acoustic model + Language model on the following spaces page (also works from browser).