What Is Predictive Coding?

Predictive coding is a data compression technique in which a signal sample is estimated from previously encoded samples, and only the difference between the prediction and the actual value is encoded and transmitted. Because adjacent samples in audio, image, and video signals are highly correlated, the prediction error is typically much smaller than the original sample value, and encoding a small error requires fewer bits than encoding the full sample. The method was first described in the 1950s in the context of speech compression and has since become a foundational building block of virtually every modern multimedia coding standard.

The technique rests on the principle that a good predictor reduces the statistical redundancy in the coded bitstream without losing information. The decoder maintains an identical copy of the predictor and regenerates the same prediction from previously decoded samples, then adds the received error to reconstruct the original signal. Because prediction and correction proceed in lockstep at both ends, no side information about the predictor output needs to be transmitted, and the only overhead is the error signal itself.

Differential Pulse-Code Modulation

Differential pulse-code modulation (DPCM) is the standard implementation of predictive coding for sampled signals. In DPCM, a prediction of the current sample is formed from a linear combination of past decoded samples, the difference between the actual sample and this prediction is quantized, and the quantized error is entropy-coded. The quantizer can be uniform or non-uniform. Stanford EE398A course materials on predictive coding describe how the predictor and quantizer together determine compression efficiency: a predictor that tightly tracks signal variations keeps the error variance small, reducing the number of bits needed by the subsequent entropy coder. A specialized case called delta modulation reduces the quantizer to two levels, encoding only the sign of each error, achieving extreme simplicity at the cost of resolution.

Adaptive Prediction

Fixed linear predictors perform poorly when signal statistics vary over time, as is common in speech and natural images. Adaptive predictive coding adjusts predictor coefficients continuously to track local statistics. In adaptive DPCM (ADPCM), both the predictor and the quantizer step size update based on the recent error history, allowing the coder to respond to changes in pitch, formant structure, or image texture. The ITU-T G.726 standard specifies ADPCM for 32 kbit/s speech coding over telephone networks and is deployed in voice-over-IP and telephony infrastructure worldwide. NASA technical documentation on enhanced DPCM covers hardware implementations of adaptive DPCM for real-time image telemetry, demonstrating compression ratios of two to four with negligible perceptual distortion.

Predictive Coding in Video and Image Standards

Predictive coding principles extend naturally to two-dimensional images and to the temporal dimension of video. In image coding, a pixel value is predicted from its neighbors in the same scan line or from the line above, and the prediction error is transformed and quantized. The JPEG-LS standard uses a context-adaptive linear predictor for lossless and near-lossless image compression, achieving higher ratios than JPEG on medical and scientific imagery. In video coding, inter-frame prediction exploits temporal redundancy by forming a prediction for the current frame from previously encoded frames with motion compensation. The overview of predictive coding from ScienceDirect notes that inter-frame predictive coding is at the core of every major video standard from H.261 through H.266 (VVC), and that the residual signal left after prediction is what the discrete cosine transform and subsequent entropy coding then compress.

Applications

Predictive coding has applications in a range of fields, including:

• Digital telephony and voice-over-IP using ADPCM standards
• Video streaming and broadcast, where inter-frame prediction underlies H.264, H.265, and AV1
• Lossless and near-lossless medical image archiving using JPEG-LS
• Satellite and remote sensing image downlinks requiring high compression at low power
• Audio coding as a component stage in codecs such as MPEG audio and FLAC