UFPAlign is an open-source automatic phonetic alignment tool for Brazilian Portuguese that uses the Kaldi toolkit for performing forced alignment of speech datasets. The UFPAlign is distributed as a plugin for Praat, a popular free software package for speech analysis in phonetics. The plugin is directly accessible from the Praat menus and it allows to align speech from an audio file and its orthographic transcription with a few minor manual steps. The result is a multi-level annotation TextGrid containing phonemes, syllables, words, phonetic and orthographic information as below.
🇧🇷 Acesse a documentação em Português Brasileiro
This assumes a installation on a
Debian-based OS like Ubuntu, so the default package manager will be apt.
Kaldi Installation
First, clone the most current version of Kaldi from GitHub by typing the following into a shell:
$ git clone https://github.com/kaldi-asr/kaldiThen, to install Kaldi tools, go to kaldi/tools/ and first check the
prerequisites for Kaldi and see if there are any system-level installations
you need to do:
$ cd kaldi/tools
$ extras/check_dependencies.shCheck the output carefully and install any prerequisites missing, like
automake, svn, and other stuff Kaldi depens on.
Then, assuming you have 4 CPU cores available on your machine, run:
$ make -j 4The last tool to install is OpenBLAS, the open-source linear algebra library that can be used instead of Intel MKL. Beware this will use all cores on your machine, even the hyperthreads if your processor has support to it.
$ extras/install_openblas.shFinally, install Kaldi src. No CUDA support is necessary.
$ cd kaldi/src
$ ./configure --shared
$ make depend -j 4
$ make -j 4To guarantee Kaldi installation was successful, run the scripts on the yes/no dataset. It doesn't take long to finish since the dataset is pretty small and the pipeline only trains and decodes a monophone-bases model.
$ cd kaldi/egs/yesno/s5
$ bash run.shThe last line should print the word error rate:
%WER 0.00 [ 0 / 232, 0 ins, 0 del, 0 sub ] exp/mono0a/decode_test_yesno/wer_10_0.0
Praat Installation
To install the Linux version of Praat, you can either use `apt-get` by typing into a shell:$ sudo apt-get install praatOr you can download a 64-bit binary executable on the Praat download page. Then, unpack it, creating the executable file praat. You can remove the tar file.
Python dependencies
Other dependencies can be installed using Python's `pip` utility:$ pip install requirements.txtUFPAlign works fine under Linux environments via command line, but also
provides a graphical interface as a plugin to Praat. In order to use the
plugin, open the New menu and click on the UFPAlign option, the following
initial window will be displayed. Click on the Choose... buttons to select
the path to Kaldi's root directory, an audio file and the corresponding
orthographic transcription. You can also choose an acoustic model, that will
be used to perform the alignment, among the five architecture available as
option. After selecting them, click on Align button. Then, wait while the
file is aligned. This may take a while.
When the alignment is successful finished, the aligner offers the option to promptly display the current resulting TextGrid in the Praat interface or to proceed to align a new audio file.
The figure at the top of this README file shows the Praat's TextGrid editor displaying an audio file waveform followed by its spectrogram and its aligner's resulting multi-tier TextGrid containing five tiers: phonemes, syllables, words, phonetic transcription and orthographic transcription, respectively.
Praat's TextGrid editor itself plots the waveform and spectrogram from the audio file, it is not content of the TextGrid file. Kaldi provides the vertical blue dashes, which correspond to the time marks, while FalaBrasil's NLP library provides the phonetic and syllabic transcriptions. Whether you decide to immediately open the result TextGrid in Praat interface or not, the TextGrid file will be saved inside a directory named textgrid at your home directory with the same name as the audio file you choose to align.
Just execute file ufpalign.sh. With no arguments, it prints a help message.
The following command works perfectly using the monophone model, though:
$ KALDI_ROOT=$HOME/kaldi bash ufpalign.sh demo/ex.wav demo/ex.txt monoFile demo/M-001.log contains an example of output from a
real execution.
Details on directory docker/execution.
In order to provide a minimum level of reproducibility of the results, the
directory simulation was created with scripts to execute MFA and other
HTK-based older aligners over the same dataset, so as to assess by comparison
the "accuracy" of UFPAlign.
We recommend the use of the docker image in docker/simulation.
If you use any of the resources provided on this repository, please cite us as the following:
Batista, C., Dias, A.L. & Neto, N. Free resources for forced phonetic alignment in Brazilian Portuguese based on Kaldi toolkit. EURASIP J. Adv. Signal Process. 2022, 11 (2022). https://doi.org/10.1186/s13634-022-00844-9
@article{Batista22a,
author = {Batista, Cassio and Dias, Ana Larissa and Neto, Nelson},
title = {Free resources for forced phonetic alignment in Brazilian Portuguese based on Kaldi toolkit},
journal = {EURASIP Journal on Advances in Signal Processing},
year = {2022},
month = {Feb},
day = {19},
volume = {2022},
number = {1},
pages = {11},
issn = {1687-6180},
doi = {10.1186/s13634-022-00844-9},
url = {https://doi.org/10.1186/s13634-022-00844-9}
}Check also the FalaBrasil's repository for training acoustic models with Kaldi.
Grupo FalaBrasil (2024) - https://ufpafalabrasil.gitlab.io/
Universidade Federal do Pará (UFPA) - https://portal.ufpa.br/
Cassio Batista - https://cassiotbatista.github.io
Larissa Dias - larissa.engcomp@gmail.com