Guide: Audio tools
ELAN is a free, easy to install media transcription program. Usually we'll use it to add transcriptions to audio files. You may want ELAN on your personal computer, so here's a link.
Praat is a free software developed by the University of Amsterdam used to display, analyze, and edit audio files. We have used it in the lab to measure voice onset times, but it has many other uses. You can also write scripts within Praat to avoid manually analyzing files.
Praat is installed on one of the RA office computers (the right-most one furthest from the doorway?). Otherwise, it is free to download here. You do not need any activation key or license.
Here is an example of what Praat looks like with an audio file open:
A few basic functions
- Open/view audio files: When Praat is open, click the dropdown tab named ‘Open’ in the leftmost ‘Praat Objects’ box. Click either ‘Read from file’ or ‘Open long sound file’ (if the file is longer than a few minutes). Select your file. The name of the file will appear in the Praat Objects list—you can add multiple files to this list so you don’t have to go back and open each one individually. Then, highlight the file in the list you wish to work with and click ‘View and Edit’.
- Listen to all or part of a file: The grey bars at the bottom of the box (see image above) tell you how long a sound or section is and allows you to play it. The lowest bar, ‘Total duration’, is the length of the entire audio file in seconds; if you click this once, Praat will play the entire audio file. If you highlight a section of the audio by dragging a box with your mouse, or click somewhere in the file, new grey bars and times will appear, aligning with the sections you’ve created. Clicking one of these once will play the audio from only that section. For instance, if you clicked once on the grey box called ‘1.343628’, it would play the audio from 0 to the red line at 1.343628 seconds. If you clicked the right grey bar, it would play from 1.34628s to the end of the file.
-
Zoom in: You can also zoom in on sounds to do more fine-grain analyses. To do this, you can either click the small box on the bottom left corner called ‘in’, then scroll to find the portion you’re looking for, OR highlight a segment of the file with your mouse and click the small box in the bottom left called ‘sel’ to zoom in on the highlighted portion. If you were to select and zoom in on the beginning of the utterance in the image above, it would look like this:
To zoom back out, click the small button that says ‘out’. - Crop: To crop an audio file, highlight the section you want to get rid of. Then, click the drop down tab, ‘Select’ at the top of the screen. Choose ‘Move start of selection to nearest zero crossing’. Then click ‘Select’ again- this time, click ‘Move end of selection to nearest zero crossing.’ Then, click ‘Edit’, then ‘Cut’. The sound file should be reduced to only the selected portion. Highlight the entire audio, click ‘File’, then ‘Extract selected sound (preserve times)’. This adds this new segment to the Object List. Then, in the list, highlight the new file, and click ‘Save’ on the top of the window, then ‘Save as .WAV file’ (usually, unless you want another file format.) You have to make sure to manually type “filename.WAV”; otherwise, it won’t save as a .WAV file. This creates a new file in the folder containing the parent file it was cropped from. If you name it with the same file name as the parent file, it will overwrite that file.
Finding the starting point for a word depends partly on the starting sound of the word. A word like 'mat' or 'ear' might start with really big, obvious waves, because the voicing comes early. Some phonemes have later voicing, like 'dog' or 'tat' or 'hard'. Here are some videos that show what that looks like when you're dealing with later voicing.
- Burst
- Plosives
- Be careful with any videos/descriptions of "Voice Onset Time", because what they're interested in is skipping the unvoiced beginning of a phoneme. we want any and all part of the sound that might precede the voicing. These two videos give examples.
- Praat tutorials Very brief overview of a lot of basic Praat functions. More in-depth tutorial There are a lot of YouTube tutorials about Praat. I found this one helpful for learning how to determine VOT.
- Praat scripts: List of Praat scripts gathered by Google.
- LCNL slack/ Debug channel: Always an option.
Py script that extracts utterance timing from an audio file. Requires a transcript, outputs timing for onset and offset at the phonetic or word level. Open-source software which requires PRAAT (below).
IBM watson keyword spotting, and Google speech-to-text are two we've used. Here's a wrapper for Watson.
Chronset (below) and Fave (above) have online versions useful for demo-ing.
Chronset is a program which detects voice onset times for short audio files, created by researchers at the Basque Center on Cognition, Brain, and Language. It requires MATLAB
There is an online platform. If preferred, you can download the source code from the same page, which needs to be run through MATLAB. An offline Chronset program is available on the middle computer in the RA room (622). Ask Mark or Calvin about getting this program on your own computer (Note: Requires MATLAB or MATLAB Runtime).
To use the online program, put up to 500 .WAV files in a ZIP folder with no subdirectory structure (i.e. no folders within the folder). Then, upload this folder to the Chronset site and enter your email address. You will eventually receive a .txt file attached to an email containing the names of your files and their VOTs (this may take up to an entire day depending on how many files you uploaded and how many are ahead of yours in the queue).
An offline version of Chronset based on the source code is available on the middle PC in the RA room (622).
- In the Start Menu, search for Chronset and open it. It may give you a warning about running it - just click "Run".
- After a few seconds, a black command prompt box will pop up and just sit there blinking while MATLAB starts up. This can take a minute or two.
- After that, a file selection window will pop up. Use this to select your files. Hold Ctrl while clicking to select multiple files. When you are ready, click "Open".
- Next, a window titled Output Method will pop up (at the same time, the command prompt window may display an error message about graphics card memory - just ignore it). The Output Method window gives you two options: "Individual" and "Batch".
- If you select "Individual", a file selection window will pop up and ask you to select an output directory; Chronset will create individual .csv files in the directory you have selected for each input file, consisting of the input file name and the calculated onset time.
- If you select "Batch", a file selection window will pop up and ask you to select an output .csv file; Chronset will create a single a single .csv file at the location you have specified with a row for each input file, consisting of the input file name and its corresponding onset time.
- After you have selected your output directory or output file, Chronset will begin processing files one at a time. The command prompt window will print out the name of the file that is currently being processed. This it slow - it may take a few minutes per file.
- After all the files are processed, Chronset will close down automatically, and your output files will be ready.
Upload specifications (Online Use): .ZIP folder; no subdirectory structure; up to 500 short audio files (around 2-3 seconds); must be a .WAV file; folder must be less than 500MB in size.
Memory Limitations (Offline Use): Calculating onset times is memory intensive, so try to keep the input files small. If you run into memory issues, try again with all other programs closed. Chronset only processes one file at a time, and writes its output as it goes along, so if Chronset crashes in the middle of processing a lot of files, you won't lose the output from before the crash.
The creators of Chronset can be reached via email at chronset@bcbl.eu.