Superflux onsets

This notebook demonstrates how to recover the Superflux onset detection algorithm of Boeck and Widmer, 2013 [1] from librosa.

This algorithm improves onset detection accuracy in the presence of vibrato.

[1]

Böck, Sebastian, and Gerhard Widmer. Maximum filter vibrato suppression for onset detection. In Proc. of the 16th Int. Conf. on Digital Audio Effects (DAFx). Maynooth, Ireland (Sept 2013), vol. 7, p. 4. 2013.

# Code source: Brian McFee
# License: ISC

We’ll need numpy and matplotlib for this example

import numpy as np
import matplotlib.pyplot as plt
from IPython.display import Audio, HTML
import librosa

The method works fine for longer signals, but the results are harder to visualize.

y, sr = librosa.loadx("trumpet", sr=44100)
HTML(librosa.util.example_info("trumpet", html=True))

solo trumpet 06

By: Mihai Sorohan | License: CC-BY-4.0

Source: https://freesound.org/s/77711/

Description

Part of "Jazz Trumpet Loops Pack in F 90 bpm" by Mihai Sorohan
All parts in this pack are in F, mostly Blues and Dorian scale and all are in stereo acidized .wav format, 16 bit, 44.1 khz sample rate, 90 bpm.

Gear used:
- trumpet Bb Amati ATR213
- microphone Oktava MK 319
- preamp Art Tube MP Project Series
- soundcard M-Audio Delta Audiophile 192

Audio(url=librosa.example("trumpet", url=True))

These parameters are taken directly from the paper

n_fft = 1024
hop_length = int(librosa.time_to_samples(1./200, sr=sr))
lag = 2
n_mels = 138
fmin = 27.5
fmax = 16000.
max_size = 3

The paper uses a log-frequency representation, but for simplicity, we’ll use a Mel spectrogram instead.

S = librosa.feature.melspectrogram(y=y, sr=sr, n_fft=n_fft,
                                   hop_length=hop_length,
                                   fmin=fmin,
                                   fmax=fmax,
                                   n_mels=n_mels)


fig, ax = plt.subplots()
librosa.display.specshow(S, vscale="dBFS[power]",
                         y_axis="mel", x_axis="time", sr=sr,
                         hop_length=hop_length, fmin=fmin, fmax=fmax, ax=ax)

Now we’ll compute the onset strength envelope and onset events using the librosa defaults.

odf_default = librosa.onset.onset_strength(y=y, sr=sr, hop_length=hop_length)
onset_default = librosa.onset.onset_detect(y=y, sr=sr, hop_length=hop_length,
                                           units="time")

And similarly with the superflux method

odf_sf = librosa.onset.onset_strength(S=librosa.power_to_db(S, ref=np.max),
                                      sr=sr,
                                      hop_length=hop_length,
                                      lag=lag, max_size=max_size)

onset_sf = librosa.onset.onset_detect(onset_envelope=odf_sf,
                                      sr=sr,
                                      hop_length=hop_length,
                                      units="time")

If you look carefully, the default onset detector (top sub-plot) has several false positives in high-vibrato regions, eg around 0.62s or 1.80s.

The superflux method (middle plot) is less susceptible to vibrato, and does not detect onset events at those points.

fig, ax = plt.subplots(nrows=3, sharex=True)

frame_time = librosa.frames_to_time(np.arange(len(odf_default)),
                                    sr=sr,
                                    hop_length=hop_length)

librosa.display.specshow(S, vscale="dBFS[power]",
                         y_axis="mel", x_axis="time", sr=sr,
                         hop_length=hop_length, fmin=fmin, fmax=fmax, ax=ax[2])
ax[2].set(xlim=[0, 5.0])


ax[0].plot(frame_time, odf_default, label="Spectral flux")
ax[0].vlines(onset_default, 0, odf_default.max(), color="r", label="Onsets")
ax[0].legend()
ax[0].label_outer()

ax[1].plot(frame_time, odf_sf, color="g", label="Superflux")
ax[1].vlines(onset_sf, 0, odf_sf.max(), color="r", label="Onsets")
ax[1].legend()
ax[0].label_outer()