Quiz

To reveal an answer, click a flashcard.

Why would we want to mark the epochs (fundamental periods) of speech, a process usually called "pitch marking"?

The primary use of pitch marks is for signal processing. That is, for manipulating speech waveforms. The most obvious example is doing that in the time domain (e.g., TD-PSOLA).

What are the advantages and disadvantages of estimating F0 from epochs (compared to using autocorrelation)?

Advantages: we can obtain both epochs and F0 in a single operation; the estimate of F0 will be precisely located in time.
Disadvantages: errors in epoch location will result in noisy estimates of F0, and deleted/inserted epochs will give octave halving/doubling errors, so some post processing will still be necessary (e.g., dynamic programming).

When estimating the spectral envelope (e.g., using STRAIGHT), we usually obtain an envelope that is too smooth - for example, formant peaks are not sharp enough. Using such a spectral envelope in synthesis can lead to "buzzy" speech. Why?

Imagine an extreme situation where the envelope becomes so smooth that it is absolutely flat, without any formant peaks. If we make a filter with that response, and excite it with a pulse train (as the source), the output will just be the pulse train itself. That will sound like, well, a pulse train: a buzzing sound. So, one perceptual consequence of an over-smooth spectral envelope is that we will perceive too much of the buzzy nature of the source.

Looking at the FFT spectrum of a frame of voiced speech, it appears to be difficult to separate out the periodic and aperiodic components because they are interleaved. What is the key insight used in STRAIGHT to estimate the amount of aperiodic energy?

The periodic component of speech is sparse in frequency (only at multiples of F0). We can assume that everything else must be aperiodic energy. Directly "removing" the harmonics would require a very precise estimate of F0, which would be error-prone. Instead, STRAIGHT assumes that the "dips" in-between harmonics are a measure of the amount aperiodic energy in that part of the spectrum.