Skip to main content

Musical chord dissonance models

Project description

This package implements various models of perceptual chord dissonance. Given a musical chord composed of individual tones, each composed of partials, a dissonance model provides a score that estimate how dissonant (harsh) does it sound to human listener.

It contains implementation of several models of dissoance with corrections of errors found in the formulas in the papers. See below.

Installation

pip install dissonant

Why dissonant? PyPI package dissonance was already taken.

Usage

Dissonance of a C major chord with harmonic tones in 12-TET tuning with base 440 Hz using the Sethares1993 model:

from dissonant import harmonic_tone, dissonance, pitch_to_freq
freqs, amps = harmonic_tone(pitch_to_freq([0, 4, 7, 12]), n_partials=10)
d = dissonance(freqs, amps, model='sethares1993')

Dissonance curve of a sliding interval of two harmonic tones: see notebooks/dissonance_curve.ipynb.

Papers

Extra

Other resources

Models

There are several acoustic models of dissonance. They may work on different data and provide several various metrics. All of them are based on some perceptual experiments.

Input signals:

  • two plain sinusoidal tones

  • two harmonic tones - integer multiples of the base frequency

  • chord (two or more) of harmonic tones

  • signals with continuous spectrum

PlompLevelt1965

They measured the perceived dissonance of pairs of sinusoidal tones and of complex tones with 6 harmonics. They provide only experimental data, not a parametric model.

Sethares1993

First, he explicitly parameterizes the data of PlompLevelt1965 (for a pair of sinusoids and for any number of complex tones) with constants found by fitting the curve to the data.

Dissonance d(x) for the difference x between a pair of frequencies. This ignores the absolute frequencies.

d(x) = exp(-a * x) - exp(-b * x)
Constants:
a = 3.5
b = 5.75

Dissonance for a pair of frequencies (f_1, f_2) (for f_1 < f_2) and their amplitudes (a_1, a_2). This takes into account the absolute frequencies. d_max is just the maximum of d(x).

d_pair(f_1, f_2, a_1, a_2) =
  s = d_max / (s_1 * f_1 + s_2)
  x = s * (f_2 - f_1)
  a_1 * a_2 * (exp(-a * x) - exp(-b * x))

Constants:
a = 3.5
b = 5.75
d_max = 0.24
s_1 = 0.0207 # 0.021 in the paper
s_2 = 18.96 # 19 in the paper

Expressed in terms of d(x):

d_pair(f_1, f_2, a_1, a_2) = a_1 * a_2 * d((f_2 - f_1) * d_max / (s_1 * f_1 + s_2))

Model of dissonance of a single complex tone (containing various partials) is just the sum of dissonances for all pairs of partials. In case the partials are integer multiples of a base frequency we can call tone “harmonic”, otherwise just a general “timbre”.

freqs = (f_1, f_2, ... f_n)
amps = (a_1, a_2, ... a_n)
d_complex(freqs, amps) = 0.5 * sum([d(f_i, f_j, a_i, a_j) for i in range(n) for j in range(n)])

or equally:

d_complex(freqs, amps) = sum([
  d(freqs[i], freqs[j], amps[i], amps[j])
  for i in range(n)
  for j in range(n)
  if i < j])

Then we can model dissonance of a pair of complex tones (timbres). Basically it’s still a sum of dissonances of all pairs of partials. We can however express freqs_2 = alpha * freqs_1 and plot d_complex() for fixed freqs_1 and varying alpha to get a “dissonance curve”.

Note that this model can be used to model dissonance of intervals of complex tones, as well as chords (any number of tones).

Note the constants s_1, s_2 are defined in the paper with low precision. Better precision is provided in the code by Mr. Sethares: http://sethares.engr.wisc.edu/comprog.html.

Questions?

Why we just sum the dissonance and not compute mean? Which aggregation operation makes more sense?

Vassilakis2001

There’s a modification to the d_pair() function which should make it depend more reliably on “SPL” and “AF-degree”, in particular put more accent on the relative amplitudes of interfering sinusoids rather than on thir absolute amplitudes.

Defined in Eq.(6.23) on page 197 (219 in the PDF).

a_2 should be >= a_1
d_pair(f_1, f_2, a_1, a_2) =
  (a_1 * a_2) ^ 0.1 * 0.5 *
  ((2 * a_2) / (a_1 + a_2)) ^ 3.11 *
  (exp(-a * s * (f_2 - f_1)) - exp(-b * s * (f_2 - f_1)))

Where:
spl = a_1 * a_2
af_degree = (2 * a_2) / (a_1 + a_2)
a = 3.5
b = 5.75
d_max = 0.24
s_1 = 0.0207
s_2 = 18.96

It can be expressed in terms of d(x):

d_pair(f_1, f_2, a_1, a_2) =
  spl = a_1 * a_2
  af_degree = (2 * a_2) / (a_1 + a_2)
  s = d_max / (s_1 * f_1 + s_2)
  x = s * (f_2 - f_1)
  spl ^ 0.1 * 0.5 * af_degree ^ 3.11 * d(x)

In comparison in the Sethares1993 model it is:

d_pair(f_1, f_2, a_1, a_2) =
  spl = a_1 * a_2
  s = d_max / (s_1 * f_1 + s_2)
  x = (f_2 - f_1) * x
  spl * d(x)

Note that in order to handle the case if a_1 < a_2 we could define:

af_degree(a_1, a_2) = (2 * min(a_1, a_2)) / (a_1 + a_2)

Looking at Vassilakis2010 this is exactly what they do, extending Vassilakis2001, otherwise the model is the same.

Extending this to a set of complex tones is the same as in Sethares1993 - just aggregate d_pair() for all pairs via a sum.

Note there’s an additional 0.5 factor in the Vassilakis2001 model compared to Sethares1993. IMHO the meaning is to compensate for pairs of partials being summed twice. In order to make the models comparable we should remove this term and take only pairs of partials only once in all models.

Cook2002

Dissonance model (Cook2002, Appendix 2, page 276 and on, eg. A2-1):

Original - wrong:
d_pair(x, a_1, a_2) =
  mu_a = (a_1 + a_2) / 2
  mu_a * (exp(-a * x) - exp(-b * x))

Fixed:
  d_pair(x, a_1, a_2) =
    mu_a = (a_1 + a_2) / 2
    mu_a * c * (exp(-a * x) - exp(-b * x))

Where:
mu_a ... mean amplitude, within [0.0; 1.0]
x ... interval between the frequencies (in semitones)
a = 1.2
b = 4.0
c = 3.5351 = 1 / (np.exp(-1.2) - np.exp(-4))

Parameters were “chosen to give a maximal dissonance value at roughly one quartertone, a dissonance of 1.00 at an interval of one semitone, and smaller values for larger intervals”.

The constants a and b are quite different than in the Sethares1993 model. Also the amplitudes are aggregated by mean, instead of product or minimum. The constant c is missing in the paper but needs to be on the place as in the formula above.

Cook2006

Other metrics: dissonance, tension, modality, tonal instability. Based on three tones, not just an interval of two.

Dissonance model (Cook2006, eq. 3).

Note that in the article it’s actually defined in a wrong way: - there’s one more minus sign when applying beta_1, beta_2 - logarithm should be of base 2 - log2(f_2 / f_1) needs to be multiplied by 12 to get the 12-TET semitone interval - there should be bracket, not floor around the difference of exponentials (probably a printing error)

Note: for the sake of readability we replace here original greek nu with v.

Original - wrong:
d_pair(f_1, f_2, a_1, a_2) =
  x = log(f_2 / f_1)
  v = a_1 * a_2
  v * beta_3 *
  floor(
    exp(-beta_1 * x ^ gamma) -
    exp(-beta_2 * x ^ gamma))

Fixed:
d_pair(f_1, f_2, a_1, a_2) =
  x = 12 * log2(f_2 / f_1)
  v = a_1 * a_2
  v * beta_3 *
  (exp(beta_1 * x ^ gamma) -
   exp(beta_2 * x ^ gamma))

Constants:
beta_1 = -0.8 # "interval of maximal dissonance"
beta_2 = -1.6 # "steepness of the fall from maximal dissonance"
beta_3 = 4.0
gamma = 1.25

Total dissonance of a chords is the sum of dissonances of all pairs of partials.

Cook2009

Simplified version of Cook2006. Still the definition is wrong (see above).

There’s no gamma exponent.

Original - wrong:
d_pair(f_1, f_2, a_1, a_2) =
  x = log(f_2 / f_1)
  v = a_1 * a_2
  v * beta_3 *
  (exp(-beta_1 * x) -
   exp(-beta_2 * x))

Fixed:
d_pair(f_1, f_2, a_1, a_2) =
  x = 12 * log2(f_2 / f_1)
  v = a_1 * a_2
  v * beta_3 *
  (exp(beta_1 * x) -
   exp(beta_2 * x))

Constants:
beta_1 = -0.8 # "interval of maximal dissonance"
beta_2 = -1.6 # "steepness of the fall from maximal dissonance"
beta_3 = 4.0

Tension of a triad (not implemented yet):

tension(f_1, f_2, f_3) =
  x = log(f_2 / f_1)
  y = log(f_3 / f_2)
  v = a_1 * a_2 * a_3
  v * exp(-((y - x) / alpha)^2)

alpha = ~0.6

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

dissonant-0.1.1.tar.gz (10.1 kB view details)

Uploaded Source

Built Distribution

dissonant-0.1.1-py2.py3-none-any.whl (15.4 kB view details)

Uploaded Python 2Python 3

File details

Details for the file dissonant-0.1.1.tar.gz.

File metadata

  • Download URL: dissonant-0.1.1.tar.gz
  • Upload date:
  • Size: 10.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No

File hashes

Hashes for dissonant-0.1.1.tar.gz
Algorithm Hash digest
SHA256 216747bd0599fabd767b6b439c5532c0461712373dfbf44e707fbdf12fe7f935
MD5 ec262a3e9fe1a858fe48b7c32a9fe252
BLAKE2b-256 a9d4f7b44ffa7a4c66188320074af28a7e2cf952bbb9e372124265197913e967

See more details on using hashes here.

File details

Details for the file dissonant-0.1.1-py2.py3-none-any.whl.

File metadata

File hashes

Hashes for dissonant-0.1.1-py2.py3-none-any.whl
Algorithm Hash digest
SHA256 e815cf882c519eb07875c7129ec8fd78b0d0e52add5c019b391285d12e03382e
MD5 367bb78ad2a940abe6496225a9ad1a2d
BLAKE2b-256 af9352409808732e7157456bdf193478a797a55c42e02425ffc48a72549e3c3f

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page