From the lectures I had long ago: to be able to encode the difference on only one bit with sufficient quality (eg low granular noise), consecutive input samples are expected to be way more correlated than differential PCM. In other words, the input source is often sampled at a higher rate in DM than in DPCM.

DPCM also supports a local predictive model (or extrapolation) not limited to zero ordre hold.

Moreover, DPCM is often followed by entropy coding to save some more bit in transmission.

From the lectures I had long ago: to be able to encode the difference on only one bit with sufficient quality (e.g. with low granular noise), consecutive input samples are expected to be way more correlated for delta modulation than for differential PCM. In other words, the input source is often sampled at a higher rate (far above Nyquist) in DM than in DPCM.

DPCM (Differential pulse-code modulation) also supports a local predictive model (or extrapolation) not limited to zero ordre hold.

Moreover, DPCM is often followed by entropy coding to save some more bit in transmission.