Module grabdata2

First stage. This does the signal processing to (essentially) compute a spectrogram of the music. The spectrogram is a two-dimensional representation of the sound: time is one axis, and the other axis is frequency.

It uses a "perceptial spectrum" which is more representative of what humans actually hear than the raw spectrogram. Essentially, it matches the spectral bandwidth to the ear: it uses a filterbank where each filter is 0.7 erb wide. Then, it half-wave rectified the signal and takes the cube root, in order to match the amplitude compression in the ear.

Parameters:

• f (str) - filename for the input audio file.
• Dt (float) - The desired sampling interval of the output feature vectors. (E.g. 10ms.)

Returns: numpy.ndarray 2-dimensional.

A sequence of feature vectors that span the specified region. This is a 2-dimensional spectrogram.

Compute beat rates from a spectrogram.

Parameters:

• sd (numpy.ndarray 2-dimensional array of floats.) - his starts from a spectrogram with one time axis and one frequency axis,

Returns: numpy.ndarray, 2-dimensional array of floats.

It computes an array where the time axis is converted to a beat-rate axis. So, the value of each element tells you how strong a particular beat is, at a particular pitch. An element might reveal the strength of a 60BPM beat in the high pitch range near 3 kHz (e.g. snare drum). So, the frequency axis separates the track by instrument, and the beats-per-minute axis lets you look at the rhythm of each instrument. (Of course, that's a somewhat idealized view...)

The basic idea here is that we should not expect a song to have a consistent rhythmic pattern thoughout. So, we divide the track into blocks, and nside each block, we assume a consistent pattern. (Obviously, this is not perfect: we chop it into blocks without regard for the song. A better scheme mght be to try to find boundaries in the song where the rhythm changes.) If the rhythm changes within a block, you'll get a blurred rhythmic pattern.

At the end, we report the average rhythm pattern, over all the blocks.

Parameters:

• f1 (str) - name of the audio file
• name (normally str) - the track's label. (Note: this is just passed through to the output.)
• plotaudio (boolean) - display plots (true) or run silently (false). This particular routine plots the averaged beat rate as a function of frequency, and also the averaged feature vector.

Returns: A pickle of (name, feature vector)

It returns an opaque blob (a pickle) that you can write to a file and easily read back. The pickle contains the track's label, along with the uncompressed feature vector. This feature vector contains the beat rate as a function of frequency, followed by the part of the feature vector that describes the averaged rhythmic pattern.