Chapter 4. Signal Processing

Front 1

What are 3 techniques for speech signal analyse?

Back 1

1. Spectral analysis
2. Cepstrum
3. Linear prediction

Front 2

What is the basic principle of Tunable bandpass and Search-tone method methods?

Back 2

The idea behind Tunable bandpass and Search-tone method methods is to use a filter (band pass, low pass) to get the parts of the spectrum you are currently interested and to suppress the rest of the spectrum.

Front 3

Explain the flow of Tunable bandpass method.

Back 3

For spectral analysis you can use a tunable bandpass which lets pass only frequencies of a small range (ΔΩ) close to the center frequency Ωm.
This results in a (complex) vibration with amplitude of the midfrequency Ωm which is proportional to the value of the spectrum.
To separate the amplitude a demodulator is used.

Front 4

What is different in Search-tone method compare to Tunable bandpass method.

Back 4

Search-tone method (windowing in the temporal domain)
Instead of changing the frequency of a bandpass filter the search-tone method uses a modulator to shift the input signal to the frequency range ∆Ω of the filter.

Front 5

What is the main problem of Tunable bandpass and Search-tone method methods and what's the solution?

Back 5

For calculating the entire spectrum at all frequencies we need to store the signal and calculate the spectrum sequentiall.
Solution:
Use bandpass filterbank with several parallel filters

Front 6

What is bandpass filter bank method about?

Back 6

Analysis all of the frequencies at the same time (with parallel filters) with absolute or relative ∆Ω
Needs windowing.

Front 7

What is the difference between different window functions?

Back 7

When selecting the shape of the window you need to consider the following criteria:
- a high trap attenuation
- a low bandpass width
The square window has a very small bandpass width but a poor trap attenuation.
The Hamming and Hann windows feature a good trap attenuation but a relatively wide bandpass.
The characteristics of the square window would lead to an imprecise and blurred spectrogram, thus Hann and Hamming windows are preferred.

Front 8

What Radix-2/Decimation-in-Time algorithm is about?

Back 8

Reduces the number of FFT needed to be counted

Front 9

What is cepstrum and what is used for?

Back 9

A cepstrum is the result of taking the Inverse Fourier transform (IFT) of the logarithm of the estimated spectrum of a signal.
Helps to differentiate noise-like and quasi-periodic sounds, helps to separate information about transfer function and fundamental frequency.
The amplitude close to zero by x-axel (quefrency) relates to transfer function of Vocal tract.
Higher amplitude by quasi-periodic signals. There are also some picks by quasi-periodic sounds which represent fundamental frequency.

Front 10

What is idea of linear prediction?

Back 10

Makes inverse filtering
Goal: separate the source signal s(t) from the filter function H(f)
Assumption: the source signal s(t) is an impulse or white noise that gets filtered by a synthesis filter
H(f) and creates the speech signal g(t) --> Source-Filter-Model

Tune the inverse analysis filter 1/H(f) until the difference between s’(t) and s(t) gets close to zero.

Use cases:
- speech compression in mobile phones
- digital coded transmission of voice through narrowband channels
- formant analysis