LAB 3: Annotation

1. In Lab 2, we calculated speech rate by using time stamped transcripts. The boundaries of the turns, however, may not be accurately labeled, as shown in the figure below. Such turn boundary errors could significantly affect your results, especially if the errors are not consistent across speakers. The purpose of this lab is to check these errors.

2. Randomly select four speakers (two female and two male) from the entire dataset (use your python code from lab1), and then randomly select 25 turns from each speaker. Write the selected turns from each speaker into a file, for example, fe_03_05838_A.txt contains the turns selected from fe_03_05838, side A. To make the Praat scripting in the next step easier (see step 5 below), you may add key words before the numbers (boundaries) of each turn, for example:

       BEG 173.54 END 175.39 A: i haven’t directly just
       ...

You’ll have four such files. Use python for this step.

3. Label the exact boundaries of the selected turns using Praat. Your TextGrids should contain a point tier, in which the initial and ending boundaries are labeled, as shown below.

4. The speech files are in the directory /corpora/Fisher/audio_wav on harris, fe_03_05838.raw_A.wav, for example, is the speech file of side A in the conversation fe_03_05838. The speech files distributed by LDC are presented in NIST sphere format, and contain two-channel mu-law sample data; and “shorten” compression has been applied to all files. The files you will be using are uncompressed (“shorten” is available on harris), and each channel goes to an individual .wav file. Praat can separate the two channels (“Read two Sounds from stereo file…”), and can recognize the NIST sphere format.

5. You need to write a Praat script to facilitate your annotation. The following procedure is recommended (Do the steps by hand, and use “paste history” to start scripting).

1. Read the file containing selected turns in to a string (Read Strings from raw text file...)
2. Read the corresponding speech file as a long sound file (Open long sound file...) [note: it’s easier to extract a portion from a long sound file in Praat]
3. Read a line of the string (e.g., turn$ = Get string... 1)
4. Find the beginning and ending times of the turn (beg = extractNumber(turn$, “BEG”), etc., see Praat string functions manual here)
5. Extract the turn from the long sound file (Extract part... 'beg' 'end' no)
6. Open a TextGrid containing a point tier(To TextGrid...)
7. Select both the extracted sound file and the TextGrid, Edit
8. Pause (for you to do the labeling)(pause)
9. Label the beginning and ending point of the turn by using waveform, spectrogram, and your ears [note: set the “dynamic range” of the spectrogram display to a suitable level (35 dB?) so that the background noise will not show up in the spectrogram]
10. Save the TextGrid into a text file (Write to short text file...)
11. Read next line of the string, and repeat the procedure

6. Write a python program to calculate the mean and standard deviation of the initial silence durations and those of the ending silence durations for each speaker. Compare the four speakers. Do you think the turn boundary errors in the original transcripts significantly affect your lab2 results? Discuss this in your lab report.