Humans are usually the intended recipients of speech signals in telecommunication, such that the quality of a transmission should be measured in terms of how good a human listener would judge its quality. Perceptual models refer to methods which try to approximate or predict the judgement of auditory quality perceived by human listeners. In coding applications we can thus define perceptual models as evaluation models, with which we approximate the perceptual effect of distortions.
An another type of models which are frequently used in speech coding are source models, which describe the inherent characteristics of the source, which is the speech signal. You can think of a source model as for example 1) physical models, which describe the physiological processes which cause speech sounds or 2) the probability distribution of speech signals. The important distinction is that source models do not care about who is observing, but they only describe the objective reality. In contrast, perceptual models are applied when we subjectively observe the signal, to evaluate properties of the signal.
In speech and audio coding applications, practically all distortions caused by the algorithms are due to quantization of the signal. The objective of perceptual modelling is then to choose the quantization accuracy such that the perceptually degrading effect of quantization is minimized. Roughly speaking, this means that those signal components which are more important to a human listener are quantized with a higher accuracy than those which are less important.
If we play two sinusoid with slightly different frequencies, then the louder of the two can mask the second sinusoid such that it becomes inaudible. This effect is known as frequency masking. In other words, people are less sensitive to sounds which are near in frequency to other sounds. In particular, when quantizing a signal, we can use a lower quantization accuracy in frequency-regions which have more energy. The effect is reduced the further away we are in frequency.
In practice, frequency masking models are similar to spectral (energy) envelopes. That is, the shape of the frequency masking model is similar to the spectral envelope, but a smoothed and less pronounced version thereof. More accurate versions of the model can be generated based on psychoacoustic theory.
Frequency masking models are used in two ways:
- In frequency-domain codecs, where a frequency-domain representation of the signal is quantized, we choose the quantization accuracy in different regions of the spectrum based on a perceptual model. Typically high-energy regions are quantized with less accuracy than low-energy regions.
- In time-domain codecs such as CELP, we typically use a analysis-by-synthesis loop, where different quantized versions are synthesized and the error between original and quantized signal is determined with perceptual weighting. The weighting is here based on a frequency masking model. Out of the different possible quantizations, the one with the smallest perceptually weighted error is chosen.
The sensitivity of human hearing depends on the frequency range where the sound is present. We are more sensitive in the "low" regions and less sensitive at "high" frequencies. There is however some ambiguity in how sensitivity is defined, and we have two prominent different interpretations:
- In the cochlea of the human ear, sounds are processed in spectral bands, which are independent such that sounds in separate bands do not interfere with each other, but sounds within the same band do interfere with the perception of each other. This is known as auditory masking. The width of these bands is frequency dependent and increases with increasing frequency. This aspect of perception has been approximated with several models, including the Bark and ERB scales.
- The distance between pitches are perceived differently depending on their frequencies. In short, a perceptually small step in pitch (measured in frequencies) is much larger at higher frequencies than at low frequencies. This aspect of perception can be approximated with the Mel scale.
The variation in accuracy and sensitivity of perception is interesting also across time. In particular, a loud sound can make imperceptible a second, weaker sound which comes later in time. Say, if we have two impulses, consecutive in time, and such that the second one is weaker, and their distance in time is sufficiently short, then we cannot hear the second impulse. Surprisingly, such temporal masking can occur also the other way around, a later loud sound can mask a preceding weaker sound.