7.1 Vowel recognition overview

Concept of vowel recognition

With this, we can proceed in one of three ways: we can model the system as having all zeros (moving average), all poles (autoregressive), or some combination of poles and zeros. Since we can only observe the output of the filter - the speech the escapes the vocal cavity - we choose to model the system as having only poles, because such a model has little dependence on the original input signal. With an autoregressive model, we can generate a transfer function to approximate the filter with a degree of precision proportional to the filter's parameter. It should be noted that a higher order model generally works better but is also more computationally expensive. Once we have the transfer function, we look at the frequency response and determine which frequencies the peaks occur. These frequencies are the formants and we can look at known formant charts to determine which vowel was spoken.

Identifying vowel formants

From the frequency response of the vowel formants, we can look at how the peaks of the frequency in the harmonic spectrum line up with the corresponding dark lines in the spectrogram. The dark areas are where the formants are and the graph show them at the same frequency.