Acoustic analysis pt. 1

Last week I finished resampling the three different age CHILDES age ranges: 0-12, 12-24, 24-36. This week I spent collecting the respective Power Spectrum Density Estimates (PSD) for each age range. For reference, here is a sketch of the process for collecting the PSD over each age range:

Read in audio files from some corpus X (i.e., childes_0-12, childes_12-24, childes_24-36, VoxForge):

AcccumulatedSamples = [];
.
.
.
[input_signal, Fs] = audioread('audiofile.wav');
AccumulatedSamples = cat(1,AccumulatedSamples,input_signal);

Accumulate samples until I have approximately 15 minutes of audio; then take the PSD and normalize:

.
.
.
if(AccumulatedSamples.length >= 900)
[Pxx, Fx] = pwelch(AccumulatedSamples(:,1),4096,[],[],Fs);
Pz = zscore(10*10log10(Pxx));

Accumulate the normalized PSD Concatenate (accumulate) the 15 minute chunks into one large matrix:

AccumulatedObs = [];
AccumulatedObs = cat(2,AccumulatedObs,Pz);
.
.
.

Because there is so much audio data and because I’m taking an FFT of each 15 minute chunk, generating the AccumulatedObs set for each corpus took a bit:

Processing Time 0-12: .81 hrs
Processing Time 12-24: 4.3 hrs
Processing Time 24-36: 4.4 hrs
Processing Time VoxForge: 2.9 hrs

Now, I’m ready to run some general statistics over the three corpuses to observe how the acoustics (spectrum) characteristics change across ages and then compare those results to the acoustic characteristics of the VoxForge data.

Best,
EO