Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/egorsmkv/flashlight-ukrainian
The Ukrainian Acoustic Model for Flashlight
https://github.com/egorsmkv/flashlight-ukrainian
flashlight speech-recognition speech-to-text
Last synced: 25 days ago
JSON representation
The Ukrainian Acoustic Model for Flashlight
- Host: GitHub
- URL: https://github.com/egorsmkv/flashlight-ukrainian
- Owner: egorsmkv
- License: apache-2.0
- Created: 2022-09-26T16:28:10.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2023-01-25T00:06:30.000Z (almost 2 years ago)
- Last Synced: 2023-03-05T22:54:55.814Z (over 1 year ago)
- Topics: flashlight, speech-recognition, speech-to-text
- Homepage:
- Size: 23.4 KB
- Stars: 7
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE.txt
Awesome Lists containing this project
README
# 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
```bash
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