What are "spectral charactersitics" of a (speech) sound?

[zaba]

Today in class, the teacher was yammering on incoherently about "spectral characteristics" but I couldn't make head or tails of it. Some kids said she was just schizo, but I think there's something to it. Can someone help me out and give me an idea of what "spectral" refers to?

[jkpate]

Acoustic signals are waves, and any wave can be decomposed into a sum of sine waves with different frequencies and amplitudes. A plot of the amplitude of each sine wave against its frequency is called a spectrum. So the spectral characteristics of a signal are patterns in the distribution of energy across frequencies. Voiceless sibilants, for example, tend to have little energy at low frequencies, with the energy gradually rising to a big hump as the frequency increases, and then falling off. Vowels, on the other hand, tend to have a few peaks in the spectrum at frequencies that end up roughly correlating with the position of the tongue in the mouth.

[zaba]

Thanks! That clears things up - but what, then, is spectral tilt?

[lx]

The effect seen on the spectrum when there is increased stress/emphasis on a specific part of the speech being examined.

[Jeff L.]

Spectral tilt is just a way to make things easier to see in a spectrogram (the visual (graph) representation of the spectrum, what you see in Praat). Spectrograms show amplitude (the "strength" of the frequency) by using darkness; if it's darker at a certain frequency, then that frequency has greater amplitude. Of course, the higher the frequency, the less energy the frequency typically has, which means it becomes harder to tell what is a formant or an increase in amplitude because it's harder to notice a more subtle difference between "rather white" and "slightly whiter". The solution is spectral tilt; the frequencies are presented as being darker (of greater amplitude) than they really are in the spectrogram so that the person analysing the spectrogram can notice the difference -- in a somewhat simplified way, you make a line following the trend of the frequencies (like you would put a line in a scatter plot to show the trend) and then you tilt the graph so that that line becomes horizontal (it's usually not quite to that extent, but it gives a more clear image). You then make the frequencies darker based on the amplitude according to this tilted graph, which means it's easier to notice the differences in amplitude in the higher frequencies.

[zaba]

Thank you very much for explaining to me how to see spectral tilt on spectrogram. While that is very helpful information, I am primarily interested in understanding what is behind that representation. In other words, what is the spectral tilt as you see it on spectrogram caused by?

[lx]

The effect seen on the spectrum when there is increased stress/emphasis on a specific part of the speech being examined.

You then make the frequencies darker based on the amplitude according to this tilted graph, which means it's easier to notice the differences in amplitude in the higher frequencies.

If those don't answer that question, then hopefully you can rephrase your question because that looks like a precise answer to me. Maybe you mean something else, but if that's the case, I think it would help to rephrase it if I've misunderstood. The increased prominence/stress/accent on a specific speech section is rendered by more power in the voice, which translates to a higher amplitude in the speech signal, which then is represented in a spectrogram by darker frequencies. You can either perceive spectral tilt like in the way I suggested, to see where increased prominence occurs, or you can use the technique like PhantomSoul said, which is to manually manipulate the amplitude to see the higher frequencies clearer (because they're often weak). Other techniques use other methods to boost high-frequency amplitudes as well, but this could be one of them.

[jkpate]

Spectral tilt does involve the tendency for energy to decrease as you go up in frequencies. However, it's not (only?) a way of drawing the spectrogram. It refers to measures of the rate at which energy decreases as you go up in frequency. People use different measures for tilt. Sometimes they take the difference between the first and second harmonic, and also between the fourth and second harmonic: a small difference means energy is dropping off slowly. Other times, they fit a linear regression to the spectrum peaks and take the slope of the regression: a slope closer to zero means energy is dropping off slowly. Still other times, they just take the difference in the total energy of a low frequency band and a high frequency band.

This phonetically interesting because shallow spectral tilt cues greater sub-glottal pressure (i.e. below the vocal folds), and this measure has been found to correlate with stress in some studies.

[Jeff L.]

This phonetically interesting because shallow spectral tilt cues greater sub-glottal pressure (i.e. below the vocal folds), and this measure has been found to correlate with stress in some studies.

...That's quite interesting -- I hadn't come across that aspect!

[zaba]

Is it accurate to say that spectral tilt is (or can be) realised as an increase in energy in the higher frequencies?

[jkpate]

It has to do with the energy in high frequencies relative to the energy in low frequencies. Simply speaking louder raises energy across all frequencies, but it doesn't necessarily increase the amount of energy in high frequencies relative to the amount of energy in low frequencies.

[zaba]

OK, let me make sure I get it.

1. So there are some acoustic differences between speaking normally and loudly  [yes!]
2. Some of those differences are spectral. [well, duh!]
3. Some of those spectral differences may be manifest as spectral tilt. [hmmm maybe]

Actually, I doubt (3) because in shouting, one does not ipso facto preserve the ratio of e in the frequency bandwidths.

Is this last sentence true?

[jkpate]

Your three points are all good Spectral tilt is just one feature of a spectrum that may be worth looking at, especially if you are interested in voice quality and prominence.

[zaba]

OK, thanks. How about for loudness like in screeching and yelling?

[lx]

Loudness is just amplitude.
If you've got full-on screaming, then the frequency component would shift up to a much higher range and there energy rate from the glottal source would be significantly higher.