Onset Detection

The goal of musical onset detection is to identify the locations in an audio signal where new sound events (such as musical notes) begin. This tutorial describes onset detection in more detail, including basics and descriptions of several algorithms.

General Use

First, an audio file must be read. We can do that using SciPy’s wavfile.

filename = 'audio_input.wav'
sampling_rate, audio = wavfile.read(file_name)
audio = np.asarray(audio, dtype=np.double)
audio /= np.max(audio)

Next, define the frame size and the hop size. These are how large each analysis window is and how large each step between windows is respectively.

This will influence:

• How many onsets you can detect in a signal of a given length, as you can detect at most len(audio) / hop_size onsets.
• The accuracy of the onset detection, as most of the detection algorithms rely on detecting changes in audio spectra over time, and larger frame sizes result in higher frequency resolution in the spectra.

frame_size = 2048
hop_size = 512

Next, compute an onset detection function:

odf = modal.ComplexODF()
odf.set_hop_size(hop_size)
odf.set_frame_size(frame_size)
odf.set_sampling_rate(sampling_rate)
odf_values = zeros(len(audio) / hop_size, dtype=double)
odf.process(x, odf_values)

odf.process(in, out) has in and out as input and output. It does not modify x, only odf_values.

Any one of

odf = modal.ComplexODF()
odf = modal.EnergyODF()
odf = modal.LPEnergyODF()
odf = modal.LPSpectralDifferenceODF()
odf = modal.LinearPredictionODF()
odf = modal.PeakODF()
odf = modal.PeakAmpDifferenceODF()
odf = modal.OnsetDetectionFunction()

can be used instead of modal.ComplexODF(). Various parameters (hop_size, frame_size) may have to be tuned.

Finally, detect onsets:

onset_det = od.OnsetDetection()
onset_det.peak_size = 3
onsets = onset_det.find_onsets(odf_values) * odf.get_hop_size()

This complete example can be found on Github.