Continuously variable slope delta modulation

Continuously variable slope delta modulation (CVSD or CVSDM) is a voice coding method. It is a delta modulation with variable step size (i.e., special case of adaptive delta modulation), first proposed by Greefkes and Riemens in 1970.

CVSD encodes at 1 bit per sample, so that audio sampled at 16 kHz is encoded at 16 kbit/s.

The encoder maintains a reference sample and a step size. Each input sample is compared to the reference sample. If the input sample is larger, the encoder emits a 1 bit and adds the step size to the reference sample. If the input sample is smaller, the encoder emits a 0 bit and subtracts the step size from the reference sample. The encoder also keeps the previous N bits of output (N = 3 or N = 4 are very common) to determine adjustments to the step size; if the previous N bits are all 1s or 0s, the step size is increased. Otherwise, the step size is decreased (usually in an exponential manner, with ${\displaystyle \tau }$ being in the range of 5 ms). The step size is adjusted for every input sample processed.

To allow for bit errors to fade out and to allow (re)synchronization to an ongoing bitstream, the output register (which keeps the reference sample) is normally realized as a leaky integrator with a time constant (${\displaystyle \tau }$) of about 1 ms.

The decoder reverses this process, starting with the reference sample, and adding or subtracting the step size according to the bit stream. The sequence of adjusted reference samples are the reconstructed waveform, and the step size is adjusted according to the same all-1s-or-0s logic as in the encoder.

...
Wikipedia