Adaptive Delta Modulation (ADM): Mechanism and Distortion Reduction

Adaptive Delta Modulation (ADM) is an improvement over classic Delta Modulation by making the quantizer step size adapt over time. Like delta modulation, it represents the signal as a running estimate updated by a small step; however, ADM increases the step size when the input changes rapidly and decreases it when the signal is relatively smooth. This dynamic step-size control is what reduces distortion—specifically slope-overload distortion and granular distortion—that occur in non-adaptive delta modulation.2

Mermaid intuition (how the step adapts to signal dynamics):

Key vocabulary for this section:

• Prediction error
• Quantization step size Δ
• Overload distortion
• Granular distortion

1) Delta Modulation baseline (what ADM fixes)

In delta modulation, the encoder compares the current input sample $x(n)$ with the current predicted estimate $\\hat{x}(n)$ and transmits a one-bit message indicating whether the estimate should move up or down by a fixed step $\\Delta$.

A common conceptual form is:

• Prediction error: $e(n)=x(n)-\\hat{x}(n)$
• Decision: transmit $b(n)=\\text{sign}(e(n))$ (i.e., “increase” if positive, “decrease” if negative)
• Estimate update:
  $\\hat{x}(n+1)=\\hat{x}(n)+b(n)\\,\\Delta$

With a fixed $\\Delta$, there are two competing failure modes:

1. Overload distortion (slope overload): if the input changes too fast, the fixed step $\\Delta$ is too small to track the slope, so the prediction error stays large and accumulates.
2. Granular distortion: if $\\Delta$ is too large relative to signal variations, the quantized updates overshoot and the estimate “roughly tracks” with visible quantization noise.2

2) Adaptive Delta Modulation (ADM): core idea

ADM modifies the above rule by allowing the step size to change with time: $\\hat{x}(n+1)=\\hat{x}(n)+b(n)\\,\\Delta(n)$

The decision $b(n)$ still depends on the sign of the prediction error, but $\\Delta(n)$ is adjusted using the recent behavior of the error/bit decisions (common practical choices: based on whether consecutive steps have the same sign, or based on an error magnitude estimate). The result is:

• When the signal keeps moving in the same direction (large, consistent error sign), ADM increases $\\Delta(n)$ to reduce overload distortion.
• When the signal direction changes or error alternates (small or oscillatory error), ADM decreases $\\Delta(n)$ to reduce granular distortion.2

Key terms in this section:

• Step-size adaptation
• Tracking capability
• Error sign sequence
• Adaptive quantizer

3) Why adaptive step size reduces distortion (mechanistic explanation)

ADM reduces distortion by changing the “effective slope” the modulator can produce.

3.1 Overload distortion reduction

For a quickly varying input (e.g., a large-slope segment of speech), delta modulation with fixed $\\Delta$ may not keep up: the maximum incremental change per sample is limited to $\\Delta$. When the input requires a larger effective increment to track its slope, the prediction error grows, which is overload distortion.

ADM addresses this by increasing $\\Delta(n)$ during periods where the error sign remains predominantly the same (meaning the estimate lags in one direction). Increasing the step size increases the maximum rate at which $\\hat{x}(n)$ can move, improving tracking and preventing error from “running away.”2

3.2 Granular distortion reduction

Granular distortion happens when $\\Delta$ is too small (or equivalently, the input is slowly varying): the one-bit decisions can cause the estimate to oscillate around the true input instead of matching it smoothly. That yields a staircase-like reconstruction with excess quantization noise.2

ADM reduces this by decreasing $\\Delta(n)$ when the bit decisions suggest that the error is alternating or becoming small—i.e., the modulator is no longer far behind and doesn’t need large corrective steps. Smaller $\\Delta(n)$ reduces overshoot and jitter, lowering granular distortion.2

Mermaid summary of overload vs granular behavior under adaptation: