Dialogue Systems	
Luděk Bártek	
Introduction to digital speech processing	
	Luděk Bártek
	Laboratory of Searching and Dialogue, Fakulty of Informatics, Masaryk University, Brno
	spring 2023
	
Sound Digitization
Dialogue Systems
Luděk Bártek
Objectives - transformation of continuous acoustic signal into a series of a discrete values, which can be processed by computer.
■ These values can be processed by computer. Sound digitization steps:
sampling - scanning the current value of the signal characteristics at given frequency (sampling frequency), quantization - transformation of values from sampling (real numbers) values into the computer real/integral number representation.
waveform coding- the way of storing information about the input signal.
Scanning the present signal value - scanning is repeating at predefined rate (sampling rate).
Sampling rate - should be double of the maximum frequency present in the signal to be able to reconstruct the original signal without loosing included information (Shannon sampling theorem)
acquired values must be quantized and stored in suitable way.
Commonly used sampling rates:
■ 8 kHz - phone quality
■ 16 kHz
■ 22050 Hz - FM quality
■ 44100 Hz - CD quality
■ 48 kHz - DVD quality
Quantization
Dialogue Systems
Luděk Bártek
Method how to transform continuous values into the
discrete ones.
Principle:
■ When the present sample cross the n-th multiple of quantization step the value n is send to the output.
■ Quantization step:
■ the average difference of input to change output by one.
■ quantization step = input values range/number of output values
■ quantization error - rounding error caused by value of quantization step, proportional to the quantization step.
Commonly used integer quantizations:
■ Sound processing: 8 bits (28 = 256 levels), 16 bits (216 = 65536 levels), 24 bits (224 = 16777216 levels)
■ image processing adds 32 bits (232 = 4294967296 levels)
Besides integer quantizations there are floating point 32 bit and 64 bit quantizations.
	Wave Coding Methods	
Dialogue		
Systems		
Luděk Bártek		
	■	Direct storing the values from previous quantization -
Introduction to digital		PCM coding (Pulse-Code Modulation).
speech processing		■ sound signal changes are relatively slow - only small
		differences of neighbouring samples.
		■ Big redundancy of data.
		■ Big amplitude dispersion problem (big quantization step
		may cause big quantization error, small quantization step
		may cause sample overflow in case of amplitude increase).
	■	Differential PCM - stores neighbouring sample differences
	■	Adaptive PCM — PCM with variable quantization step
		value - quantization step adapts to the signal amplitude.
		
Dialogue Systems
Luděk Bártek
Introduction to digital speech processing
Based on following assumptions:
■ Difference of neighbouring samples is much less than the value of a sample.
■ The following sample can be relatively precisely estimated as a previous samples linear combination.
Differential PCM coding block schema
5(11)
1
s"<n)
s'(n)
s"(n) - sample assumption
s'(n) - reconstructed signal, gained from the following sum quantized signal 5'(n) a s"(n) S(n) = s(n) - s"(n)
Adaptive Pulse Code Modulation
Dialogue Systems
Luděk Bártek
Big signal amplitude changes may cause:
■ Inaccurate week signal capture - amplitude is too small, comparable to quantization step (quantization step is too big).
■ Strong signal distortion - overflow of a range of values used to signal coding (quantization step is too small).
Solution: adapting quantization step to a signal amplitude.