How to prepare dataset for training

Download Ukrainian dataset from https://github.com/egorsmkv/speech-recognition-uk.
Delete Common Voice folder in dataset
Download import_ukrainian.py and put into DeepSpeech/bin folder.
Run import script
Download Common Voice 6.1 Ukrainian dataset
Convert to DeepSpeech format
Merge train.csv from dataset and from DeepSpeech into one file
Put CV files into dataset files folder
Put dev.csv and test.csv into folder

Note: you can also specify dataset with "," e.g. dataset1/train.csv,dataset2/train.csv.

You have a reproducible dataset!

Scorer

Refer to DeepSpeech guide for further explanations.
Generate scorer package.

python3 generate_lm.py --input_txt ../../../voice-recognition-ua/data/all_text.txt --output_dir . \
  --top_k 500000 --kenlm_bins ../../../voice-recognition-ua/kenlm/build/bin \
  --arpa_order 5 --max_arpa_memory "85%" --arpa_prune "0|0|1" \
  --binary_a_bits 255 --binary_q_bits 8 --binary_type trie

Run lm_optimizer to find the best scorer value.
Rerun step 2 to generate new scorer.

Caution: scorer is very model-dependant, so you'll likely need to adjust it to each model.