The Ukrainian Acoustic Model for Flashlight

🇺🇦 Join Ukrainian Speech Recognition Community - https://t.me/speech_recognition_uk

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

Overview

This repository contains the acoustic model for Ukrainian trained on Flashlight framework: https://github.com/flashlight/flashlight/tree/main/flashlight/app/asr

• Architecture: Conformer (30m params)
• Data in train: Common Voice 10 & Voice of America
• Trained epochs: 410
• Train time: around a week

Quality

• WER: 9.07775% (id est the quality is 90.92%)
• TER: 1.98391%

Download

All files are here: https://github.com/egorsmkv/flashlight-ukrainian/releases/tag/v1.0

How to test?

Run a container with Flashlight running with CPU

docker-compose up

# and in another termianl
docker exec -it flashlight_cpu bash

Run

Just with an AM:

/root/flashlight/build/bin/asr/fl_asr_test --am /models/uk_am.bin --datadir ''  --emission_dir '' --uselexicon false \
 --test /data/rows.lst --tokens /models/tokens.txt --lexicon /models/lexicon.txt --show

With a LM:

/root/flashlight/build/bin/asr/fl_asr_decode \
 --am=/models/uk_am.bin \
 --test=/data/rows.lst \
 --maxload=3477 \
 --nthread_decoder=2 \
 --show \
 --showletters \
 --lexicon=/models/lexicon.txt \
 --uselexicon=false \
 --lm=/models/lm_4gram_500k.binary \
 --lmtype=kenlm \
 --decodertype=wrd \
 --beamsize=200 \
 --beamsizetoken=200 \
 --beamthreshold=20 \
 --lmweight=0.75 \
 --wordscore=0 \
 --eosscore=0 \
 --silscore=0 \
 --unkscore=0 \
 --smearing=max

• rows.lst is from https://github.com/egorsmkv/cv10-uk-testset-clean
• lm_4gram_500k.binary is from https://huggingface.co/Yehor/kenlm-ukrainian/tree/main/news/lm-4gram-500k

How to fine-tune on own data?

/root/flashlight/build/bin/asr/fl_asr_train continue /models/ --flagsfile /models/train.flags

/models/ must contain .bin files