This repo is currently a documentation for installation and execution of mfa
The official installation of MFA depends on conda, hence here is an alternative virtualenvironment version of it. Beware, the mfa needs openfst, openblas, ngram, baumwelch and sox which I am not installing here.
virtualenv --python=python3 venv
pip install montreal-forced-aligner
It will give a warning about a conflicting version of numpy, hence:
pip uninstall numpy
pip install numpy==1.20.3
and finally:
pip install montreal-forced-aligner -U
pip install pynini
mfa thirdparty download
First download the models for the desired language:
mfa download acoustic turkish
mfa download g2p turkish_g2p
generate the lexicon from the g2p models and the text file.
mfa g2p turkish_g2p judeo-esp/corpus/wav/article-1.txt judeo-esp/lexicon.txt --clean --temp_directory ./tmp/
Finally align the audio with the transcription
mfa align judeo-esp/corpus/ judeo-esp/lexicon.txt turkish . --clean --temp_directory tmp
ffmpeg -i judeo-esp/Lo-Ke-Mos-Aze-Ganar-I-Perder-El-Koronavirus-Seli-Gaon.mp3 -ac 1 -ar 16000 judeo-esp/corpus/wav/article-1.wav
sed 's/sh/ş/g; s/dj/c/g; s/\(.*\)/\L\1/' judeo-esp/article-1-org.txt > judeo-esp/corpus/wav/article-1.txt
mfa g2p turkish_g2p judeo-esp/corpus/wav/article-1.txt judeo-esp/lexicon.txt --clean --temp_directory ./tmp/
mfa align judeo-esp/corpus/ judeo-esp/lexicon.txt turkish . --clean --temp_directory tmp