Characterising non-linear behaviour of coupling capacitors through audio feature analysis and machine learning

Clarke CJ, B T B, Chen JM. Characterising non-linear behaviour of coupling capacitors through audio feature analysis and machine learning. In: AES Convention 150. Audio Engineering Society; 2021. Paper 10463. Available from: https://aes.org/publications/elibrary-page/?id=21056

@inproceedings{Clarke2021_21056,
  author = {Clarke, Christopher Johann and B T, Balamurali and Chen, Jer-Ming},
  title = {{Characterising non-linear behaviour of coupling capacitors through audio feature analysis and machine learning}},
  booktitle = {AES Convention 150},
  note = {Paper 10463},
  year = {2021},
  month = may,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=21056}
}

TY  - CPAPER
TI  - Characterising non-linear behaviour of coupling capacitors through audio feature analysis and machine learning
AU  - Clarke, Christopher Johann
AU  - B T, Balamurali
AU  - Chen, Jer-Ming
T2  - AES Convention 150
M1  - Paper 10463
PY  - 2021
DA  - 2021/05/06
UR  - https://aes.org/publications/elibrary-page/?id=21056
PB  - Audio Engineering Society
LA  - en
AB  - Different electrically-equivalent capacitors are known to impact the sonic signature of the audio circuit. In this study, the non-linear behaviour of five different coupling capacitors of equivalent capacitance (marketed as "audio capacitors"), one at a time, are characterised. A dataset containing the input and output signals of a non-linear amplifier is logged, its audio features are extracted and the non-linear behaviour is analysed. Machine learning is then applied on the dataset to supplement analysis of the Total Harmonic Distortion (THD). The five capacitors’ THD performance seem to fall into two categories: below 200 Hz, there is significant standard deviation of 14.1 dBc; above 200 Hz, the capacitors show somewhat similar behaviour, with only 0.01 dBc standard deviation. This separation however, does not hold at regions below 0.2 V. A support vector machine model is trained and classifies the five capacitors well above chance: the best classification at 84% and worst at 36%. The methodology introduced here may also be used to meaningfully assess the complicated behaviour of other audio electronic components.
ER  -