Prak was presented at the International Congress of Phonetic Sciences 2023. See article pre-print in arXiv and Prak ICPhS poster.
View slides from the Prak presentation at FňUK/FÚ Nov 8th 2022 (or download the slides here).
OK. Get the Prak source code and python3.10 with pytorch and torchaudio. Record a wav, make a phrase tier with text and prak will find the phone and word times:
prak -i pocasi.textgrid -w pocasi.wav -o pocasi.out.textgrid
Enjoy!
Given a Czech voice recording and corresponding text transcript, this tool can guess a sequence of phones and their position in time.
Design goals:
- Fully open-source, avoiding dependencies which would restrict academic and/or commercial use in any way. Original code was written for this project from scratch to avoid any legal limitations and allow others to continue this work.
- Compatible with Linux, Mac, and Windows.
- More precise than competing tools, specifically targeting the needs of Czech phoneticians.
Go to mambaforge, download mambaforge for windows (Mambaforge-Windows-x86_64). Run it, despite any protests from Windows (it will whine about installing an unknown application, shrug it off through 'More information' and then 'Run anyway'). Install to C:\mambaforge. This location is the only thing you have to change during the installation process, you can go with the default settings in the other installation dialog windows.
Run a mambaforge console. It should be in your Start menu under the name "Miniforge Console" (the window will be called "Miniforge Prompt"). The text prompt should start with (base). (If it doesn't, you opened the wrong executable. Find the right one.) Type this (and hit return):
mamba install pytorch torchaudio -c pytorch
...and then this (then return again):
pip install pysoundfile
If any Y/N questions pop up, just hit return (or type Y and then return if you feel like it).
Download the Prak ZIP file (also available through the green "Code" button at the top right of this GitHub page → Download ZIP.). Uncompress it directly to C:\ (or move it there after uncompressing it elsewhere), creating the C:\prak-main folder. Rename this folder to C:\prak.
Run Praat. Open the script C:\prak\prak_align_phrase.praat and add it to the menu (File > Add to dynamic menu > Class1: Sound, Class2: TextGrid, Command: Align using Prak).
Set Praat's encoding to UTF-8 (Praat → Preferences → Text reading preferences → UTF-8, then Text writing preferences → UTF-8).
Refer to Prak installation details if you need to know more.
Go to mambaforge. Choose and download the installation file here, either Mambaforge-MacOSX-x86_64 for older Intel-based Macs or Mambaforge-MacOSX-arm64 for new Macs with Apple M1 chips. Then make the downloaded file executable (using chmod) and run it in the terminal, for example (for Intel Mac):
chmod +x Downloads/Mambaforge-MacOSX-x86_64.sh
Downloads/Mambaforge-MacOSX-x86_64.sh
Answer a few questions (likely yes or just Enter but you must explicitly type yes where needed). If the licence agreement is too long to show, press spacebar untill you reach the end, then agree (yes). After installing mambaforge successfully, QUIT TERMINAL COMPLETELY and run it again. The prompt will now start with "(base)" and you can install python packages we need:
mamba install pytorch torchaudio -c pytorch
Get prak, preferably using git (makes future updates super easy):
git clone https://github.com/vaclavhanzl/prak.git
cd prak
(Or you can get zip as described above for Windows and uncompress it in your home folder.)
Run praat. Open script C:\prak\prak_align_phrase.praat and add it to menu (File > Add to dynamic menu > Class1: Sound, Class2: TextGrid, Command: Align using prak). Set read/write to UTF-8 (Praat > Preferences > Text reading preferences > UTF-8, then Text writing preferences > UTF-8).
Refer to Prak installation details if you need to know more.
Things happen too fast in the python world for apt package managers to keep up. So you most likely want a special package manager just for python, and there are really great tools to choose from.
For scientific work, conda package manager might be better than pip. In the conda world, there are still many options. You likely do not want the huge Anaconda but rather the more free and modular conda forge. To get it working, you still have multiple options from which mambaforge (faster conda) looks quite good. With this general guidance, it is now easy to google your way.
Big part of pytorch installation complexity stems from the CUDA GPU drivers installation. If you do not plan training big neural networks or do not have a decent GPU, you may very well use pytorch just on the CPU. Prak only uses CPU for phone alignment and even acoustic model can be reasonably trained on just the CPU.
You need pytorch and torchaudio packages. As a linux hacker you likely have other things to do with python so prak tries to use its own virtual environment prak before resorting to base or even trying to run without a virtual environment
(this logic is in prak/prak). Prepare it like this:
mamba create prak
mamba activate prak
mamba install pytorch torchaudio -c pytorch
Get prak source code via git. If you have a GitHub account and ssh key, clone directly from your home like this:
git clone git@github.com:vaclavhanzl/prak.git
cd prak
otherwise resort to https:
git clone https://github.com/vaclavhanzl/prak.git
cd prak
Then run praat and add prak/prak_align_phrase.praat to menu, the same way as described above for Mac.
If you did not put mamba or prak directly to your home directory, edit config variables in this script.
Refer to Prak installation details for even more details.
Select sound and textgrid containing phrase tier, press your new Align using prak button. You can also align multiple sounds using multiple textgrids, or multiple sounds using one input textgrid for all. Some details:
- text can be in one or more intervals of tier named "phrase" or "Phrase"
- if input textgrid does not contain tier with such name, prak will nicely let you choose the right one
- times of the text tier do not have to match the audio so you can align several wavs using one textgrid
- text can contain all the usual interpunction (if it is human readable, prak will likely process it, too)
- text is expected to be utf-8 NFC but prak can deal with problems like BOM, NFD or CRLF
- prak automatically matches selected wav(s) and textgrid(s) and either goes on automatically if it makes clear sense or asks you for details if needed
- if the textgrid contains a phone tier, prak warns you before going on to safeguard any manual tuning you already did
- sound can have any sampling frequency
- several pronunciation variants are considered where there is possible ambiguity and the (hopefully) right one is selected based on acoustic signal match
- you can specify your own exceptions for foreign words
- phone tier created by prak uses SAMPA phone codes
Most of the phone-phone boundaries will likely still need manual microtununig to suit your needs but not much more than that. On the level of words (and silences present or not present between words), alignment should be mostly correct. Prak tries its best to decide whether voicing assimilation happened or not on word boundaries, whether glottal stop is present, whether some possible phone changes happened or not etc. Often it is right, sometimes the decision is hard even for a human. Sometimes prak gets these things wrong but it should be a minority of cases.
For alignment with audio, you will need a Czech voice recording (wav) and a corresponding Czech transcript (in phrase tier in an utf8 TextGrid).
The transcript can contain usual punctuation (as a secretary would transcribe it) but should be precise even at
disfluencies - if the recording contains a repeated word, so should the transcript. If you used praat to
make a phrase tier with transcript and saved it to file pocasi.textgrid, you can use prak to guess
pronunciations and time align all phones and words. Run this in terminal:
~/prak/prak.py -i pocasi.textgrid -w pocasi.wav -o pocasi.out.textgrid
The Align using prak button is in fact using this method internally to do its magic.
To get an idea about pronunciation variants being considered, run this first (being in the prak folder):
prongen/prak_prongen.py
The tool will nicely help you. To see generated Czech pronunciation, try this:
prongen/prak_prongen.py -p -s
or, if you are on Mac or Linux or you somehow got a real terminal in Windows, try much nicer
prongen/prak_prongen.py -p
and then type Czech sentences in terminal (finish each with Enter, press Ctrl+D to stop). (Look at prongen/README.md for details.) If this works, python is installed OK.
Multiple pronunciation variants are considered, the acoustic model will hopefully choose the right
one. You may need to teach the tool some additional foreign words or tell it about important seams in composite
words. Just edit the prak/exceptions.txt file.
It is automatically used at this default location by the GUI method and you can also use it from command line
using the -e option.
If you used git to download prak, it is easy. Go to the prak folder and run:
git pull
If you went with the zip option, download a new zip the same way.
You certainly want to update, prak is getting better often! See recent list of published changes or even what happens in the develop branch (extremely courios ones can even see feature branches).
If you feel very conservative, you can use zips from Releases (like 1.0.0) - these are older but more tested.
1 minute audio was phone-aligned in 18 seconds on a 2014 Mac Air, in 3.6 seconds on a 2020 M1 Mac Air, in 5 seconds on a decent 2020 Intel linux box.
The acoustic model was trained solely on the Czech CommonVoice recordings. I am also immensly grateful to Fonetický ústav FFUK for the possibility to compare their hand labeled (and phone-level time aligned) data with prak's automatic labeling, this helps a lot. However, training of the prak's acoustic model currently does not involve any hand-aligned data. (CommonVoice data consist of recorded sentences with corresponding ortographic transcriprion. No human-decided time alignment on the level of words or phones is used for training, these alignments are purely machine made.)
Some Czech phonetitians call similar tools "nastřelovač" as these tools position phones and their time stamps in a fast but rather unprecise manner, like positioning objects by shooting them where they should be. The Czech word "prak" means "sling", a common shooting toy.
Of yourse! You are welcome! You can contribute in many ways:
- Try to install and use prak and report any issues, e.g. in Discussions. Use version from the develop branch to help with tests of the latest changes.
- Become a developer
- Add pronunciations of foreign words or help for harder Czech words, like composed Czech words with vowel after seam.
You can discuss prak in public here in Discussions. If you want to tell me more personally that you love this or hate this, message Vaclav Hanzl on Mastodon.
I'd like to thank all the nice people from Fonetický ústav FFUK and equally nice young people from FňUK who inspired me by their charming approach to science and life.
Look at Prak wiki to learn more details than you ever wanted to know.