Audio Features and Vamp Plugins

General outline

To be propagated to / synchronised with

  1. Introductory notes and slides on acoustics and audio (CR, 25 min)
  2. Introductory notes and slides on audio features (CC, 5 min)
  3. Sonic Visualiser - hands on with waveform and spectrograms (CC, 25 min)
  4. Introductory notes and slides on Vamp plugins (CC, 5 min)
  5. Sonic Visualiser - hands on with Vamp plugins (CC, 15 min)
  6. Python/IPython intro (CR, 35 min)
  7. Break (20 min)
  8. Feature Extraction using Vamp Plugins in Python (IPython Notebook) (CC, 40 min)
  9. Audio Indexing and Search in Python (IPython Notebook) (CR, 40 min)

Breakdown of CC sections

Direct links to the audio files themselves are included here, but for copyright reasons they won't necessarily persist beyond the workshop!

Sonic Visualiser - hands on with waveform and spectrograms
  • Waveform
    1. Start Sonic Visualiser and open A Friendly Warning (very severe, synthetic 80s pop song by Act)
    2. Show dragging through the file using Navigate tool, and also using the overview at bottom
    3. Play from the start, just to get an idea what it sounds like
    4. Return to the start and zoom in (using the zoom wheel, but noting that the mouse wheel also works)
    5. Notice the different shapes in waveform resulting from different types of synthetic percussive sound (low-frequency kick drum / higher frequency cymbal-type sounds) - refer back to Christophe's notes about correspondence between e.g. signal voltage and speaker cone deflection
    6. Continue until the vocal starts, and observe that we can see very little that relates to e.g. sung pitch, although if we zoom in we can quite clearly see sibilance (these frequencies around 10kHz are pretty much the sweet spot for visibility in a 44.1kHz waveform)
  • Spectrogram
    1. New session, open piano-scale.wav and play it
    2. Some information can sort-of be perceived and measured from the waveform here: we can see when the notes start, and can get simple fundamental frequency estimate - zoom in to the first note, switch to Select mode, drag out one cycle - it's about 170 samples, so 44100/170 = 259 Hz - the note is a middle C so true value should be nearer to 261, but this is a fair approximation. (But this is a very simple example!)
    3. Now open a plain spectrogram - Pane -> Add Spectrogram (or G key). Observe full range on frequency scale; x axis is time, this is a simple time-frequency breakdown.
    4. Notice that, for each note, we can see the fundamental frequency most strongly and then the harmonic sequence. The harmonics are spaced more widely for higher notes because they are multiples of the fundamental frequency, which is larger. The noise floor is visible because we're using a dB scale, can switch to Linear to isolate only the strong frequencies. There isn't really enough detail to measure much here. (NB the default colour scheme is unhelpful to colour blind users, so it might be worth changing to Sunset scheme.)
    5. Close that pane and open a "melodic-range spectrogram" - Pane -> Add Melodic Range Spectrogram (or M key). Observe the much more limited frequency range and the fact that this spectrogram uses both Linear colour and Sunset scheme by default.
    6. Although the higher harmonics quickly disappear off the top of the scale, we can clearly see that the spacing between harmonics is now the same for each note, but the harmonics for a given note get closer together as they go up; and that the semitones are equally-spaced (note spacing corresponding to the major scale intervals). This is because the vertical scale is now logarithmic in frequency, which makes it (fudging the issue a little) linear in pitch. Correspondingly there is now a little representation of a piano keyboard shown at left, with middle C highlighted.
    7. The above assumes 12tET with A=440Hz, we can change at least the latter part of that in the Preferences and the scale will move immediately when we do so (demonstrate but be sure to restore the default)
    8. Select the Measure tool and show that we can get a frequency readout with harmonic markers. Return to the Navigate tool and contrast with the readout that is displayed as you move the pointer over the pane.
    9. Close that pane and open the "peak-frequency spectrogram" - Pane -> Add Peak Frequency Spectrogram (or K key). Notice that here we can just wave the Navigate tool over a bin to get an estimate of the instantaneous frequency there.
    10. New session, open A Friendly Warning again and open both the plain spectrogram and the melodic-range one -- observe and contrast the various visible elements, in particular vertical lines in full frequency range for noisy percussion, relative invisibility of such broadband sounds in the melodic-range spectrogram, glides in vocal, difficulty of distinguishing harmonic traces from simultaneous notes etc.
    11. Go to File -> Replace Main Audio, open King Henry. Note among other things that we need to increase the gain on the melodic-range spectrogram, and that the vibrato is visible and things like vibrato rate could be approximately measured, that the long reverb makes the notes appear to overlap in places.
Introductory notes and slides on audio features
Sonic Visualiser - hands on with Vamp plugins
  • Quick survey of nominally high-ish level feature extractors
    1. Start a new session with A Friendly Warning.
    2. Feature extraction plugins are found under the Transform menu, so called because it contains things that turn your audio into something else, including both feature extractors and audio effects. You can run a transform and show the output in the same pane as the audio, or in a new one.
    3. With the audio pane selected, run Transform -> Analysis by Category -> Time -> Tempo -> Bar and Beat Tracker: Beats. This produces a series of beat locations, each labelled with metrical beat number, and when you play the audio, the beats are played with a tap.
    4. Now open a new pane (Pane -> Add New Pane or shortcut N) and run Transform -> Analysis by Category -> Time -> Onsets -> Note Onset Detector: Onsets. Here we have individual note onset positions: it works OK for this sort of music. We can switch playback on and off for the individual feature tracks with the Play toggle on the parameter box at right.
    5. Some further transforms we can try: Chordino chord estimate; the notes corresponding to Chordino's chord estimate (so as to audition whether the chords sound any good); QM key estimator and key-strength plot.
  • Pitch
    1. Close session, return to King Henry. Open a melodic-range spectrogram and make sure its pane is current.
    2. Now run the transform pYIN - Smoothed Pitch Track. A pitch track should appear in a bright colour. Switch its plot type to Discrete Curves and make sure its scale is set to Auto-Align, which means that if it has Hz units, it will be aligned to the same vertical scale as the spectrogram behind it.
    3. We can of course check this extracted pitch-track visually, but we can also inspect individual values (by mouseover), inspect in bulk (Layer -> Edit Layer Data) including tracking through the data during playback, and export to a file (File -> Export Annotation Layer). Demonstrate this latter. (Note that the correct layer must be selected for any of these to work!)
    4. This layer can also be synthesised and played back -- switch on the Play button on layer parameters and try it.
    5. The same plugin can produce note segmentations (for monophonic audio of this type), so run it again requesting the Notes output. Each note is recorded as having a pitch equal to the median of the underlying pitch track's pitches for the time it spans. This is unlikely to sound so nice when played back, because of both segmentation flaws and difficulties (e.g. glides) and interesting properties of pitch perception (e.g. with vibrato). If this kind of use is of interest to you, consider our other program Tony.
  • Chroma reduction
    1. Using the NNLS Chroma plugin - open an empty pane and run this transform with the default parameters. This is a single-octave reduction of frequency content (can explain at arbitrary length).
    2. Run the same transform again in a second pane, but this time with different parameters: local tuning, L2 norm, spectral shape = 0.9.
Feature Extraction using Vamp Plugins in Python

Following notebook.

Feedback from test audience on SV/Vamp section (not Python)

  • Make more of a point of telling us what we're about to see, before we see it
    • On which subject: pull the "audio features" introductory slides back to the start (before any SV) and make the connection between "spectral features" on slide and "spectrograms" in SV
  • Drop the last Vamp Plugin slide, it's just too confusing
  • Add a quick pic (or live view) of the Vamp plugin download page
  • General feeling that we should talk about "feature extractors" more and "plugins" less, especially since they are not classified plugin-by-plugin in SV's user interface -- "plugin" seems a bit of a developer term and why is it called "Vamp" anyway
  • It's not clear that having audience be hands-on with SV is helpful given the time available -- it would be, if they had additional time to carry out some work on their own recordings, but as it is it might just be a distraction
  • Drop the bit on the peak-frequency spectrogram, or else figure out how to explain it without having introduced the concept of a spectrogram "bin" already because that is surprisingly difficult to just handwave about