I am looking to find out how accents could/would be detected by a programme, I haven’t as of yet undergone an experiment into detecting or observing anything noticeable about recorded accents. I am not sure a frequency/harmonic analyses will reveal anything. Trying to find any literature review on the subject is sadly lacking, so I was hoping to find out from the community where one might start or what would you look for in any given accent. Something quantifiable could or should surely be observed dependant on an accent?

For instance say I have 10 people from town A and 10 people from town B, all males with roughly the same age Town A and Town B both speak English but have different accents. If I ask each of them to repeat the same word, we should notice a difference in the frequency/harmonic analysis of Town B folk from Town A folk.

What is your best guess at where this could be found? Is frequency or harmonic analysis going to reveal anything between the two groups?

For instance if we take specific vowels we could possibly see a certain threshold reached in amplitude level and frequency (pitch) and could then work out the resonance frequencies of the vocal tract for certain accents based on key vowels? In speech science and phonetics the use of formant is used to mean an acoustic resonance of the human vocal tract. It is often measured as an amplitude peak in the frequency spectrum of the sound, using a spectrogram or a spectrum analyzer, though in vowels spoken with a high fundamental frequency, as in a female or child voice, the frequency of the resonance may lie between the widely-spread harmonics and hence no peak is visible.

If you notice that using a formant table you can construct vowels with a vocal formant filter, could this be reverse engineered to essentially "detect"?

My worst fear however is that any given vowel is going to have either a lack of accent specific information, or the analysis done will have too much sporadic information between individuals.

Could anyone point me in the right direction here im quite stuck?

  • 1
    $\begingroup$ It seems that for pitch or some other feature of sound the sound of a word could represent a short time series for the individuals. If the sample is not too short you could perhaps fit harmonic regression models for individuals and look to see if amplitude and frequency vary more between members from different towns than for comparing members from different towns. $\endgroup$ – Michael R. Chernick Sep 19 '12 at 12:44

I would suggest creating a dataset of voice recordings, labelled by accent. This will allow you (eventually), to create a model that can distinguish between accents. Obviously the more speech samples you have per subject and per accent, the more likely you will be able to perform this task reliably.

After appropriate filtering, you could then break each recording into (domain relevant) chunks or epochs (e.g. 10s epochs) and then calculate a number of standard speech signal processing or standard signal processing features for each epoch.

In terms of the signal properties I think accent would manifest itself in terms of both amplitude and frequency characteristics. Some features to try per epoch:

  • Mean amplitude per epoch
  • RMS amplitude per epoch
  • Amplitude range per epoch
  • Mel-frequency cepstral coefficients (MFCCs)
  • Spectral edge frequency (calculated from windowed power spectral density per epoch)
  • Harmonic frequency ratios (literature search may be necessary here to determine frequency ranges)
  • Power in relevant frequency ranges (again, application specific)

Plotting the features for each accent group might be a good first pass for determining which features can discriminate. If you have more than one subject you could take the mean and coefficient of variability of each feature across all epochs to allow comparison between subjects. An ANOVA could then serve as a good starting point to test for significance difference between different accent groups for a given feature.

Hope this helps.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.