LAB 4: F0 declination

1. F0 tends to decline over the course of phrases and utterances, both in non-tonal and tonal languages. This global downward trend of F0 is called declination. In this lab, we’ll explore some interesting questions related to declination: Do we start with a higher F0 when producing longer sentences? Do longer sentences have a flatter declination slope of F0? Studies of such questions can lead us to a better understanding of pre-planning of pitch contours, and pre-planning of speech production in general.

2. We'll use the TIMIT corpus, which contains a total of 6300 sentences, 10 sentences spoken by each of 630 speakers from 8 major dialect regions of the United States. There is a copy of the corpus on harris: /corpora/timit.

3. Extract the F0s of the 6300 sentences, and write the data (time-F0 pairs) of each sentence into a text file (using Praat). Then extract the following variables for each sentence: the initial F0 of the sentence, the end F0 of the sentence, the slope of the regression line over the F0 curve of the sentence, the intercept of the regression line, the length (duration) of the sentence, and the total number of words in the sentence; write the extracted variables from all the sentences into one file for further analysis (using Python).

4. Since speakers have very different pitch ranges, we need to normalize the F0 data for each speaker before extracting the variables. Below is the (suggested) procedure for normalization and extracting variables:

   For each of the 630 speakers:

1. Pool the F0 data of the 10 sentences of the speaker;
2. Find the Q1 (25 percentile) and Q3 (75 percentile) points of the data, and calculate the Whiskers: Q1-1.5*(Q3-Q1) and Q3+1.5*(Q3-Q1);
3. Find the minimum and maximum F0 values within the Whiskers;

For each of the 10 sentences of the speaker:

       1. Check the transcript file of the sentence, if the end punctuation mark of the sentence is not a period (e.g., a question mark), or if there is a puncuation mark in the middle of the sentence, discard this sentence (why?), and go to next one. Otherwise, record the total number of words in the sentence;
       2. Read in the F0 file, set the time of the first F0 to be 0, and then adjust the times of the other F0s;
       3. If an F0 value is outside the whiskers (a potential outlier), remove it (and its time) from the data; otherwise, normalize the F0 using the following formula: F0_norm = (10/log(maxf0/minf0))*log(F0/minf0), which maps the F0s onto a log scale of 0-10.
       4. Record the initial and end F0_norm and their times;
       5. Do a linear regression on the F0_norm data (F0_norm ~ time), record the slope and the intercept of the regression line (you can use the llinregress function in the stats package);
       6. Write the following into your report file (one file for all speakers):
          speaker_ID; sentence_ID; total_words; sentence_length; initial_F0_norm; end_F0_norm; end_F0_norm - initial_F0_norm; slope; intercept.

5. Do regression and correlation tests to find whether initial F0, end F0, the difference between initial and end F0s, slope, or intercept is correlated with sentence length or the total number of words in the sentence. If interested, you can further look at if there are any differences between male and female, or among different dialect regions (using the speaker information file included in the corpus).

6. Your lab report should (at least) include exploratory data analysis on each of the variables (histograms, box-plots, five-number summaries, etc.); regression and correlation results (coefficients, p values, scattlerplots with regression line etc.); and a brief discussion on your results.