Gammatone Filter Initialization Techniques

• Gammatone filter initialization is the process of configuring filter coefficients, impulse shape, and phase to replicate key characteristics of human auditory filters.
• The approach leverages psychoacoustic scales, such as ERB and Cam, to determine center frequencies and bandwidths that align with biological and computational models.
• Practical implementations incorporate methods like instance normalization and convolutional architectures to optimize performance in speech recognition and auditory modeling.

Gammatone filter initialization refers to the theoretical formulation and practical methodology for specifying the coefficients, parameters, and spatial/frequency arrangement of the gammatone family of auditory filters in digital or analog signal processing systems. These filterbanks are central to computational auditory modeling, psychoacoustic analysis, end-to-end audio neural architectures, and biologically-inspired front-ends. Their initialization—encompassing impulse shape, frequency coverage, quality factor, and phase—controls system behavior prior to adaptation through learning or further optimization.

1. Mathematical Formulation of Gammatone Filters

The canonical gammatone filter impulse response is defined as

$$g(t) = A \cdot t^{n-1} \cdot e^{-2\pi b t} \cdot \cos(2\pi f_c t + \phi), \quad t \geq 0,$$

where

• $A$ is amplitude (gain normalization),
• $n$ is the filter order (number of poles in cascade or the exponent),
• $b$ is the bandwidth (envelope decay constant, Hz),
• $f_c$ is center frequency (Hz),
• $\phi$ is a phase offset (rad).

Multiple generalizations exist:

• The generalized gammatone uses cascaded integrators with potentially non-uniform time constants, yielding richer spectral and temporal characteristics (Lindeberg et al., 2014).
• Discrete-time implementations may use IIR cascades (commonly “Slaney’s method”), mapping continuous decay and frequency to digital pole placement and gain (Isoyama et al., 2023, Lindeberg et al., 2014).

2. Selection of Center Frequencies and Bandwidths

Selecting the grid of center frequencies $f_c$ and corresponding filter bandwidths is typically guided by psychoacoustic scales that mimic the human auditory system’s resolution:

• ERB (Equivalent Rectangular Bandwidth) scale: $ERB(f) = 24.7 (4.37f/1000 + 1)$ Hz (Isoyama et al., 2023, Lin, 2017).
• ERB-number (“Cam”) scale: $Cam(f) = 21.4 \log_{10}(1 + 4.37f/1000)$; invert for $f = \frac{1000}{4.37} (10^{Cam/21.4} - 1)$ (Isoyama et al., 2023).
• Logarithmic scaling: $\Upsilon(f) = 7.7\,\ln(f) - 23.1$; bandwidths $f_B = f_c / 7.7$ (“log ERB”) or $24.7 + 0.108 f_c$ (“linear ERB”) (Lin, 2017).
• Uniform coverage: Ensure subbands are equidistant on $\Upsilon$ and that coverage metric $$\eta_C^{(b)} = (\frac{1}{2}(f_B^{(b)} + f_B^{(b+1)})) / (f_C^{(b+1)} - f_C^{(b)})$$ is constant (Lin, 2017).

The number of channels and endpoints are chosen to span the signal domain (e.g., 20–8000 Hz for speech) with sufficient overlap (typically $\eta_C \gtrsim 1$).

3. Parameter Estimation and Psychoacoustic Constraints

Historically, gammatone filter constants were chosen based on simultaneous masking data [Slaney 1993]. Recent studies have introduced characteristics-based frameworks:

• Magnitudinal characteristics: Peak frequency, $n$ dB bandwidth, ERB, Q-factor, convexity.
• Phase/group delay: $\tau_{\text{peak}} = g/(2\pi p_z CF)$, with $g$ the order and $p_z$ a normalized bandwidth constant (Alkhairy, 1 Jan 2026).

Empirically validated parameterizations include:

• Classical setting: $g = 4$, $p_z(CF) = 0.0252 \cdot (4.37\,CF + 1)/CF$ (CF in kHz)
• Updated physiological setting: $g \approx 7.2$, $p_z(CF) = 0.0354 \cdot (4.37\,CF + 1)/CF$ or $0.1303 \cdot CF^{-0.27}$

Initialization thus requires selecting order $g$, computing $p_z$ for each $CF$, then $b_m = p_z \cdot CF_m$, and phase $\phi = -(\pi/2) g$, fixing amplitude so that $|H(f_c)| = 1$ (often $A = (2\pi b)^n / \Gamma(n)$) (Alkhairy, 1 Jan 2026).

4. Practical Initialization Algorithms and Implementations

Table: Core Initialization Steps and Variants

Source	Center Frequency Placement	Bandwidth	Order	Phase	Notes
(Zeghidour et al., 2018)	Mel/ERB scale (ref impl.)	ERB-based	4	not stated	Neural front-end; instance norm crucial
(Isoyama et al., 2023)	Uniform ERB-number (Cam)	ERB-based	4	$\phi=0$	ISO 532-2 loudness, cascade IIR, IIR norm
(Lin, 2017)	Equidistant on log scale	log/linear	4	not stated	Consistent frequency coverage; Q factor
(Lindeberg et al., 2014)	Specified	Specified	4–5	$0$ or fitted	Recursive 1st order, uniform/log-distr.
(Ditter et al., 2019)	1 ERB steps, 100–4000 Hz	ERB-based	2	multiple	Multi-phase, truncated, low-latency
(Alkhairy, 1 Jan 2026)	Specified/any	Characteristic-based	$4$ or $7.2$	$-(\pi/2)g$	Modern psychoacoustic fit

Additional steps may include truncating impulse responses for latency (Ditter et al., 2019); normalizing in frequency or amplitude domain; and, for convolutional architectures, loading coefficients into the appropriate tensors (Ditter et al., 2019, Zeghidour et al., 2018).

5. Architectural Modifications and Their Effects

Several modifications to the front-end affect the criticality and utility of initialization:

• Instance normalization: Essential for stabilizing and accelerating training in deep models; found to eliminate the performance gap between random and gammatone-based initialization in end-to-end speech recognition (Zeghidour et al., 2018).
• Low-pass filter choice: Replacing max-pooling with a fixed squared Hanning window further desensitizes learning outcomes to gammatone initialization, allowing random seeds to perform equivalently and simplifying the deployment process (Zeghidour et al., 2018).
• Convolutional architectures: In modern neural front-ends, gammatone filterbanks are instantiated as convolution kernels and may be further optimized by backpropagation. However, deterministic gammatone or multi-phase gammatone filters can yield strong performance without further parameter learning, and ablation studies confirm comparable or improved metrics relative to learned baselines (Ditter et al., 2019).

6. Initialization for Biological Validity and Engineering Constraints

The scale-space theory justifies generalized gammatone filters, allowing additional tuning between frequency selectivity and latency. For modeling early auditory structures:

• Order $K=4$–5 and log-distributed poles achieve sub-10 ms group delay as observed in biologically recorded receptive fields (Lindeberg et al., 2014).
• Window length $n \approx 4$ yields filter ERBs that match mammalian inferior colliculus and cortical responses.

Recent psychoacoustic estimates support higher filter order ($g \geq 7$) and revised bandwidth constants to align with auditory nerve fiber tuning, providing a route for initializing filters that accurately reflect both behavioral and physiological sharpness (Alkhairy, 1 Jan 2026).

7. Applications and Impact in Computational Audition

Gammatone filter initialization underpins the following domains:

• End-to-end neural acoustic models: Trainable or deterministic filterbanks used as front-ends for speech recognition and separation; initialization choices influence convergence speed, robustness, and final error rates (Zeghidour et al., 2018, Ditter et al., 2019).
• Psychoacoustic and sound-quality metrics: Banks aligned to ERB and Cam scales support direct computation of metrics such as loudness (per ISO 532-2), sharpness, and fluctuation strength with minimal RMSE relative to human data (Isoyama et al., 2023).
• Auditory neuroscience modeling: Filter parameters and arrangements reflective of recent biological findings yield feature extraction models that reproduce both tuning curves and group delay seen in ICC and A1 (Lindeberg et al., 2014, Alkhairy, 1 Jan 2026).
• Low-latency speech processing: Truncated multi-phase gammatone banks enable sub-2 ms latency in causal stream separation, outperforming learned approaches in generalization and robustness (Ditter et al., 2019).

The trend in recent research is toward psychoacoustically updated, characteristics-based filter initialization, moving beyond legacy parameters toward empirically validated, transparent, and biologically meaningful auditory models (Alkhairy, 1 Jan 2026).