xarxaxdev/montreal_forced_aligner_tachelhit

Master Thesis Project for Potsdam university: MFA is a method to generate timestamped transcriptions for a language, given only the audio (and using many pre-transcribed text and audio as training). # montreal_forced_aligner_tachelhit Master Thesis Project for Potsdam university: MFA is a method to generate timestamped transcriptions for a language, given only the audio (and using many pre-transcribed text and audio as training). # Run to setup the environment ``` conda create -n aligner -c conda-forge python=3.11 montreal-forced-aligner pip -y # as advised by MFA docu # just enter until its completely created conda activate aligner # All these are run within the new environment export THREADS=14 # change for whatever you like conda config --set default_threads $THREADS conda env config vars set OMP_NUM_THREADS$THREADS conda env config vars set OPENBLAS_NUM_THREADS=$THREADS conda env config vars set MKL_NUM_THREADS=$THREADS conda deactivate && conda activate aligner pip install datasets==3.6.0 pip install soundfile==0.13.1 pip install torch==2.10.0 pip install torchaudio==2.10 pip install torchcodec==0.10 ``` # Prepare data for the aligner ``` conda activate aligner # Activate your environment python prepare_paths.py # cleanup previous files; generate needed paths # We need to train individual languages in the order # kab > zgh > tzm/shi # Generate unified vocabulary and all pronunciation dictionaries python kab_build_dicts.py python zgh_build_dicts.py # we will consider shi/tzm as one python merge_dicts.py # generate all.dict python kab_gen_corpus_acoustic_model.py # CAREFUL; THIS IS COMPUTATIONALLY HEAVY python zgh_gen_corpus_acoustic_model.py python shi_gen_corpus_acoustic_model.py python tzm_gen_corpus_acoustic_model.py ``` # Train aligners ``` alias mfa_train='mfa train --clean --single_speaker -j 12 --overwrite' alias mfa_adapt='mfa adapt --clean --single_speaker -j 12 --overwrite' alias mfa_align='mfa align --clean --single_speaker -j 12 --overwrite' # Heavy computational work. Beware. # mfa train [OPTIONS] CORPUS_DIRECTORY DICTIONARY_PATH OUTPUT_MODEL_PATH # 1 job = 1 core, I am using 14 here # --single_speaker is required to par …