Lab #3 LPC Formant Tracking
S702 D. Kewley-Port W. SU 2/26/02
Ferguson and Kewley-Port (2001) reported that the second formant for high front vowels was higher for a male speaking in the "clear" speech style than in a "conversational" style. New data has been collected from 41 additional talkers. You will measure the F2 in /i/ in two word pairs from two male talkers to determine whether they also have higher F2 values for their /i/ vowels.
The purpose of this lab is to introduce you to Colea, a set of MatLab routines written by Philip Loizou: http://www.utdallas.edu/~loizou/speech/colea.htm
Colea is very powerful, with many LPC, pitch tracking and other speech analysis tools. Most likely you will not explore all of them in this lab, but you are encouraged to try lots of them. LPC does not perfectly track formants, particularly in some consonant environments. You will need to explore LPC options to determine if the formant tracker is giving you the correct values for the syllables you will analyze. All analysis packages have their limitations. In this case, the formant tracker in Colea fixes parameters like the order M, the window size N etc. Only the LPC Spectra window associated with the cursor allows you to change these parameters (in the Control window) and to compare LPC and FFT spectra. If you want to change M in the Colea formant tracker, you would need to reprogram one or more of the .m files in the Colea package! However, there are other ways you can explore to determine the correct formant values (e.g. selecting a different smaller piece of the waveform to analyze) and maybe the convenience of Colea can work for you.
The steps in this lab are:
1) Create and put all your files in one directory on
your zip drive.
Start with Colea website and downloading "colea.zip", and upzip it.
Remember! in order to unzip the file, you need to "Extract" the files by choosing the "open it" when you click on the file name, and the Winzip window will come out. Click on "Extract" to save under your zip disk (removable disk).
2) Download eewavs.zip to your zip drive (same directory) from Oncourse->group files->Data. Unzip the 8 files for this lab. There are two talkers (MO3 and MX2), clear (CL) and conversational tokens (CON), and /i/ in /bid/ and in a diphthongized context, "tire" in CL and "wire" in CON.
3) Start MatLab. Set the directory to your zip drive directory. Type colea in the command window. Load file M03beadCON.wav as your test example. Remember to "clear all" everytime when you pull up a new .wav file, because Colea might crush or "go crazy" if you don't type "clear all".
4) Become familiar with the LPC and formant tracking features of Colea. Wen will show you the procedures in the Lab section. There is also "manual.pdf" that you have unzipped that may be helpful.
5) Now you are ready to measure F2 for the /i/ vowel in each of the 8 waveforms. Enter the values in a small table. Calculate an average of the F2 for /i/ for the two CL and CON tokens (seperately for each talker). Determine if either one or both of the talkers had a higher F2 in CL than in CON speech. In your report, relate your results (albeit minimal) to those of Ferguson and Kewley-Port.
6) Under the menu File, "load and stack" the two waveforms of "MO3bead.CL" and "MO3bead. CON" for clear and conversational tokens. By listening and looking at the waveforms, observe and report on whether talker MO3 made other differences between the his clear and conversational tokens of /bid/.
7) For at least one of the 8 files, Colea badly mistracks the formants. Print out the formant tracks for such a file. Note on the print out what you preceived the problem was, and what solutions you made to get correct formant values. For your report, make any general comments you think appropriate about using Colea for formant tracking.
REFERENCE: Ferguson, S. H. and Kewley-Port, D. (2002)
Vowel intelligibility in clear and conversational speech for
normal-hearing and hearing-impaired