Psychoacoustics and Perceptual Coding
The development of efficient audio compression algorithms depends critically on understanding how the human auditory system processes sound and what aspects of audio signals are perceptually important. Psychoacoustic research has revealed numerous phenomena that audio codecs exploit to reduce data requirements while maintaining perceptual fidelity.
The critical band structure of human hearing divides the audible frequency range into approximately 24 overlapping bands, each corresponding to a specific region of the basilar membrane in the inner ear. Audio signals within each critical band interact strongly, while signals in different bands can be processed somewhat independently. This property enables audio codecs to analyze and quantize each critical band according to its specific perceptual requirements.
Critical band bandwidths increase with frequency according to empirical formulas:
CBW(f) = 25 + 75[1 + 1.4(f/1000)²]^0.69 Hz
Where f is the center frequency in Hz. This relationship shows that critical bands are narrow at low frequencies (about 100 Hz wide below 1 kHz) and become progressively wider at high frequencies (several kHz wide above 10 kHz).
Simultaneous masking occurs when loud sounds make quieter sounds inaudible if they occur at the same time and fall within the same critical band. The masking effect is strongest for frequencies close to the masker and decreases for frequencies farther away. Masking curves can be modeled as:
M(f) = M₀ + S × log₁₀(f/fm)
Where M₀ is the masking threshold at the masker frequency fm, and S is the slope of the masking curve. These curves enable audio codecs to calculate which signal components can be eliminated without perceptual consequence.
Temporal masking extends the masking effect across time, with loud sounds making quieter sounds inaudible both before (pre-masking) and after (post-masking) their occurrence. Pre-masking lasts about 5-20 milliseconds, while post-masking can extend up to 100-200 milliseconds depending on the masker characteristics.
The absolute threshold of hearing establishes the minimum sound levels audible in quiet conditions across the frequency spectrum. This threshold exhibits a complex frequency dependence with maximum sensitivity around 3-4 kHz and reduced sensitivity at very low and very high frequencies:
T_quiet(f) = 3.64(f/1000)^(-0.8) - 6.5e^(-0.6(f/1000-3.3)²) + 10^(-3)(f/1000)⁴ dB SPL
Audio codecs can eliminate any signal components that fall below this absolute threshold, as they would be inaudible even in perfect listening conditions.
Spectral and temporal resolution limitations of human hearing enable further compression opportunities. The auditory system cannot detect rapid changes in spectral content or precise timing of events, allowing audio codecs to use time-frequency analysis windows that trade temporal resolution for spectral resolution in perceptually optimal ways.
The Modified Discrete Cosine Transform (MDCT) used in many audio codecs provides this time-frequency analysis:
X(k) = Σ[n=0 to N-1] x(n) × cos[π/N(n + 1/2 + N/2)(k + 1/2)]
Where x(n) is the input signal and X(k) are the transform coefficients. The MDCT provides good frequency resolution for analyzing masking relationships while maintaining reasonable temporal resolution for handling transient events.
Perceptual entropy represents the theoretical minimum bit rate required to encode audio signals without perceptual loss, calculated based on psychoacoustic masking models. This metric provides a target for codec designers and helps evaluate the efficiency of different compression algorithms.
Advanced psychoacoustic models account for additional perceptual phenomena: - Binaural masking effects in stereo signals - Complex interactions between multiple maskers - Individual variations in hearing sensitivity - Listening environment and reproduction system characteristics
These sophisticated models enable next-generation audio codecs to achieve higher compression ratios while maintaining or improving perceptual quality.