Dithering is a technique commonly used to improve the perceptual quality of lossy data compression. In this work, motivated by low-complexity wearable applications, we analytically and experimentally justify the use of dithering for Automatic Speech Recognition (ASR) input compression, evaluated on OpenAI's Whisper Model. We hypothesize a relationship between the mean squared error, the autocorrelation vector of the quantization error evaluated at a specific lag$(\tau=5)$, and ASR performance under lossy compression. Using this hypothesis, we propose a parametric dithering technique for a low-complexity speech compression pipeline. Specifically, we define two dithers indexed by$m$as:$f_{V_{m, \alpha}}(v)=\alpha \Lambda_{2 \Delta((\alpha-1) m+2-\alpha)}(v)+(1-\alpha) \delta(v), \quad m \in \{1,2\}$where$\Lambda_{2 a}(v) \triangleq \frac{1}{a^{2}}(a-\vert v\vert)$for$\vert v\vert \leq a$, with$\Delta$representing the quantization interval size and$\alpha$as the control parameter.

Dithering is a technique that can improve human perception of low-resolution data by reducing quantization artifacts. We hypothesize that the perceptual prominence of quantization artifacts is proportional to the magnitude of the quantization error autocorrelation vector. Under this hypothesis we derive two parametric dither distributions that trade-off between minimizing mean square error and minimizing an upper bound on the quantization error autocorrelation vector magnitude in the ℓ 1 sense $\left( {{f_{{V_{1,\alpha }}}}(v) = \alpha {\Pi _{\alpha \Delta }}(v) + (1 - \alpha )\frac{1}{2}\left[ {\delta \left( {v - \frac{{\alpha \Delta }}{2}} \right) + \delta \left( {v + \frac{{\alpha \Delta }}{2}} \right)} \right]} \right)$ or ℓ 2 sense $\left( {{f_{{V_{2,\alpha }}}}(v) = {\Pi _{\alpha \Delta }}(v)} \right)$ where ${\Pi _a}(v) \triangleq \frac{1}{a}, - \frac{a}{2} \leq v \leq \frac{a}{2}$ and ∆ is the width of the quantization region. The application of these distortion-controlling dithers to an example low-rate image recompression problem (using Lena) reveals optimal performance with partial dithering (0 < α ∝ λ < 1) as per Fig. 1 while our novel ℓ 1 -optimized dither produces a new Pareto front for the quality-entropy trade-off shown in Fig. 2 .