AlignmentDuration

Tool for Aligning lyrics to audio automatically using a phonetic recognizer with Hidden Markov Models. The Viterbi Decoding with explicit durations of reference syllables can be toggled on with the parameter WITH_DURATIONS

Built from scratch. Alternatively one can use this tool as a wrapper around htk (may be faster) by setting the parameter DECODE_WITH_HTK

If you are using this work please cite http://mtg.upf.edu/node/3751

NOTE: A version building upon this research is built by Voice Magix. It features

• latest deep-learning enabled acoustic model
• English language lyrics parser and normalizer
• runtime speed optimization
• option to run on recordings with diverse types of background instruments
• reduced external package dependencies

If interested in using it write to info at voicemagix dot com

Folder Structure

• example: example/test sound and annotation files
• scripts: help scripts for running the code (including on hpc cluster )
• src: main source code
  • align: main alignment logic
  • hmm: hidden Markov model alignment
  • for_makam: Makam-specific logic (see music traditions below)
  • models_makam: acoustic model for Turkish
  • models_jingju: acoustic model for Jingju Mandarin
  • for_jingju: jingju-specific logic (see music traditions below)
  • onsets: logic for note-onset-aware alignment (ISMIR 2016)
  • parse: logic for parsing lyrics files
  • smstools: modifications to the https://github.com/MTG/sms-tools
  • utilsLyrics: any utility scripts
• test: test scripts (scould be used in CI)
• thrash: code that has to be reviewed and deleted, left for the sake of completeness

LICENSE

Copyright 2014-2017 Music Technology Group - Universitat Pompeu Fabra

AlignmentDuration is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation (FSF), either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/

For more details see COPYING.txt

BUILD INSTRUCTIONS

NOTE: python3 is not supported and tested

git clone https://github.com/georgid/AlignmentDuration.git; sudo apt-get install python-dev python-setuptools python-numpy

pip install -r requirements; python setup.py install

• essentia.

• pdnn needed if OBS_MODEL is MLP or MLP_fuzzy

cd ..; git clone https://github.com/yajiemiao/pdnn

install also Theano

• install htk needed if either MFCC_HTK or DECODE_WITH_HTK is set to 1

• htkModelParser needed if on Turkish makam and OBS_MODEL is GMM

git clone https://github.com/georgid/htkModelParser.git; cd htkModelParser; sudo pip install -r requirements; python setup.py install

• sci-kit-learn (branch with fixed underflow issues) needed if on Turkish makam and OBS_MODEL is GMM

git clone https://github.com/georgid/scikit-learn; sudo apt-get install python-scipy; python setup.py install

• git clone https://github.com/georgid/makam_acapella needed if using MLP_fuzzy model

• Evaluation (optional if evaluation of accuracy needed) <put in parent directory of AlignmentDuration>

cd ..; git clone https://github.com/georgid/AlignmentEvaluation.git

Citation

Georgi Dzhambazov, Knowledge-based Probabilistic Modeling for Tracking Lyrics in Music Audio Signals, PhD thesis thesis materials companion page

USAGE on different music traditions

jingju (Beijing Opera) : Chinese

python AlignmentDuration/jingju/runWithParamsAll.py 2 0 /JingjuSingingAnnotation-master/lyrics2audio/results/3folds/ 3 0

to test: python AlignmentDuration/test/testLyricsAlign.py

with method testLyricsAlign_mandarin_pop

Turkish Makam music: Turkish

You need to provide the musicbrainz ID (MBID) of the recording. This requirement could be removed on demand...

call as a method from an aggregator API:

install https://github.com/MTG/pycompmusic; python pycompmusic/compmusic/extractors/makam/lyricsalign.py

or locally:

python https://github.com/georgid/AlignmentDuration/blob/noteOnsets/src/for_makam/lyricalign_local.py

to test: python AlignmentDuration/test/testLyricsAlign.py

with method testLyricsAlignMakam

English

Write to georgi.dzhambazov at upf dot edu or info at voicemagix dot com if you would like to use the English language model. It is not included here for licensing issues.

Evaluation

Use evalAccuracy script. 100 means perfect alignment. Usually values above 80% are acceptably well for human listeners.

The default evaluation level is set at word boundaries

BUILD INSTRUCTIONS ON server kora.s.upf.edu

git clone https://github.com/georgid/AlignmentDuration.git git checkout for_pycompmusic

cd /homedtic/georgid/test2/AlignmentDuration source /homedtic/georgid/env/bin/activate python setup.py install

to test: python /homedtic/georgid/test2/AlignmentDuration/test/testLyricsAlign.py

on server: git pull https://github.com/MTG/pycompmusic /srv/dunya/env/src/pycompmusic/compmusic/extractors/makam/lyricsalign.py with recording MB-ID: 727cff89-392f-4d15-926d-63b2697d7f3f