Plot your vowel space

The purpose of this assignment is for you to gain familiarity with connected speech and automated speech processing tools, and ultimately to investigate your own acoustic vowel space. You will record yourself reading a story and then use an automatic transcription tool to segment the recorded speech, then use another automated tool to measure the formants, and finally upload your formant file in order to generate a vowel plot that will be e-mailed to you automatically.

  1. Get praat from Download and install the correct version for the computer you are using. Some university computers already have it installed.

  2. Get the files. Download get_formants.praat, which will be used to measure formants, and OHDARE2.txt, a story used by Erik Thomas. If the files open in your browser when you click them, try right-clicking and slecting "Save link as...". If that doesn't work, try another browser, or copy the text to text files on your computer and save them with the correct names.

  3. Make a test recording. Open Praat's SoundRecorder (New...Record mono Sound...). Set Sampling Frequency to 22,050 or 44,100. Click "Record", then talk into the microphone (the colored bar should light up when you talk), then click "Stop", enter a name in the "Name:" field, and click "Save to list" (Avoid using any spaces in the filename. Use "_" instead.). Now when you go back to the main Praat window, there will be a Sound object with the name you gave it. (If you prefer, you can make a recording using Audacity, or CoolEdit, or any other sound recording program, save it as a .wav file, and then open it in Praat (Read...Read from file.).

  4. Open your sound in the editor. Highlight your sound in the Object list, and click "Edit". Inspect and listen to your recording. If it sounds distorted, try recording again with the record level set lower. If it sounds weak with a lot of background noise, try recording again with the record level set higher.

  5. Record the story. Go to a computer where you can make a recording and Practice reading the story a few times. Then record it. If you make a mistake while reading, you can repeat the part where you made a mistake, and then cut out the mistake in the next step.

  6. Inspect your recording Praat. Inspect and listen to your recording using Praat. Use the Editor to remove any extra time before or after the reading of the story, or any parts you had to repeat (by selecting the parts you want to remove and pressing Ctrl-X (or Command-X). Save the sound as a .wav file by selecting the sound in the object list and choosing Save... Save as WAV file

  7. Segment your file automatically. Use a version of the Penn Phonetics Lab Forced Aligner that is running on the Phonetics Lab server at NCSU to align a transcription (instead of doing it manually).

  8. Check the alignment. Open the story .wav file and the TextGrid you received by e-mail in Praat. Select them and click "Edit". Look at the spectrogram and the TextGrid tiers and see if the aligner made any mistakes. The Forced Aligner uses a plain text version of IPA transcriptions. Click on the word and phone labels in the TextGrid tiers and click below to play them. Make sure they really sound like what they say. If you find minor problems, you can drag the boundaries or change the labels and then save the TextGrid file.

  9. Adjust the parameters for Praat's formant tracking algorithm. In the editor window, select Formant...Show formants (if it's not already selected). You should see rows of red dots superimposed on the formants in the spectrogram. Praat needs some help knowing where to look for the formants. Depending on the pitch of your voice, you may need to adjust a parameter. Select Formant...Formant settings. For an adult cisgender female, a good starting point is a "Maximum formant (Hz):" value of 5500. For an adult cisgender male, a good starting point is value at 5000. For a small child's voice, a starting point of 6000 is realistic. These are suggestions, and the best way to find an actual good max formant value for your speech is to start with what seems like a realistic max formant value and then look at the formant tracking and make adjustments higher or lower as needed. Now look at the formant tracking to see if the red dots mostly follow the formants in the spectrogram. Look for vowels that you expect to have extreme F1 or F2 values: "UW1" should have low F1 and low F2, "IY1" should have low F1 and high F2, and "AE1" should have high F1. If the formants aren't tracking well, change the "Maximum formant (Hz):" value until they look better.

  10. Record the formant values for all of the words. In the "Praat Objects" window, click Praat...Open Praat script..., and then click on get_formants.praat. Click Run...Run. In the window that pops up, change the value for "sound file" to the name of your Sound and TextGrid (they should have the same name if you ignore the ".wav" and "TextGrid" parts). Change "maximum formant" to the "Maximum formant (Hz):" that worked best in the previous step. Leave "number of formants" at 5, unless you have a reason to change it. Click "OK". The script may take a few minutes, but when it is done, you should see a file in the same folder as your other files called formants_[your name].txt. It contains measurements from the three points in each vowel. You can open it up and inspect it in a text editor or spreadsheet program. You should see one line for each vowel, with the word it occurred in and the segments next to it, followed by six formant values.

  11. Submit your file. Upload the formant file (the text file that was just created in the folder where all your other homework files are, not one of the objects from the Praat object list) and enter your e-mail address. When you see "Done. E-mail sent.", check your e-mail. The e-mail you receive will contain the vowel plot as a pdf file and information about how good your vowel plot looks, and your current mark for the assignment.

  12. File name:
    E-mail address:
  13. Revise and resubmit. Chances are there will be some problems with your first vowel plot. If the ellipses are huge, you probably have a lot of formants that didn't get measured properly. Find vowels that are in the wrong place, and then look on another page of the file to see what words those were in. Then go back to Praat and find those words in the Editor by clicking in the "word" tier and then click Edit...Find... and enter the word you are looking for (the search is case-sensitive). (1) Check to see if the vowel was segmented properly: If you play the segment it should sound like the vowel. If not, move the boundaries. (2) Check to see if the formants were tracked properly: The red dots should track the formants. If they are in the wrong place, try changing the formant settings as above. (3) If there is a problem you can't solve (e.g., you mispronounced or didn't pronounce the word), you can just delete the segment label in the "phone" tier, and it will be ignored. (4) Save your TextGrid if you made any changes. Then repeat the last three steps until the plot looks good.