Typical ADPCM codecs are lossy because prediction differences are quantized to a limited set of levels.

Does adaptive mean the bitrate changes continuously?

No. A codec can adapt its internal predictor or quantizer while operating at a fixed selected bitrate.

Can any ADPCM decoder open any ADPCM recording?

No. Variant, sample format, packing, initial state and container details can affect compatibility.

AI Glossary / MLOps and deployment / Intermediate

Adaptive differential pulse-code modulation

Definition

Adaptive differential pulse-code modulation (ADPCM) is a family of audio-coding methods that represents the difference between a sample and its predicted value, while adapting coding parameters to the signal. It usually reduces the number of bits needed per sample by accepting some reconstruction error.

What it means

Ordinary sample coding describes an amplitude directly. Differential coding instead asks how far the actual amplitude lies from a prediction. If neighboring samples are predictable, their differences can require fewer bits than the full amplitudes.

The adaptive part adjusts the quantization scale, the predictor, or both, depending on the particular codec. Quiet, slowly changing material and rapidly changing material need different treatment. ADPCM is a family name, so two products described as ADPCM are not necessarily compatible.

This is signal compression, not a machine-learning model. Its adaptation can follow fixed mathematical rules without a training dataset. It also differs from lossless differential encoding: quantizing the difference generally prevents exact recovery of the original samples.

How it works

A predictor estimates the next sample from information already available to the decoder. The encoder subtracts that prediction from the actual input to obtain a difference, also called a prediction error.

A quantizer maps the difference to one of a limited number of representable levels. The transmitted code identifies that level. The encoder also reconstructs the sample locally by adding the quantized difference to the prediction. Using reconstructed history helps keep its internal state aligned with the decoder, which cannot see the original input.

The decoder performs the corresponding reconstruction and state updates. Adaptive rules change the quantizer scale or prediction behavior as the signal evolves. Larger steps accommodate larger differences but provide coarser resolution; smaller steps offer finer resolution within a narrower range.

ITU-T G.726 is one standardized example, supporting 16, 24, 32 and 40 kbit/s operation. Its exact arithmetic, state and packing rules determine interoperability.

Three practical examples

1. Voice transport: A communication system can use an agreed ADPCM codec to carry speech with fewer payload bits than an eight-bit-per-sample telephone PCM stream.

2. Embedded sound storage: A device with limited storage may keep spoken prompts in an ADPCM format, provided its decoder supports that exact variant and the resulting quality meets the product's needs.

3. AI audio ingestion: A transcription service may receive ADPCM recordings from an existing device. It must decode them correctly before preparing the waveform for speech recognition; renaming the file as PCM will not perform the conversion.

Think of it this way

Think of reporting a vehicle's movement by saying how far it has moved from the last agreed position instead of sending the whole position every time. Small movements can be described economically, and the reporting scale can change when movement becomes larger.

The analogy has a limit: ADPCM predicts samples using codec-specific state, and quantized differences introduce approximation. A decoder cannot silently substitute its own prediction or rounding rules and expect the same reconstructed signal.

Worked example

Consider an invented differential coder whose predictor is simply the previous reconstructed sample. Its initial reconstructed value is 100, and differences are rounded to the nearest multiple of 4.

The next actual sample is 113. Prediction is 100, so the difference is 13. The coder represents this as 12 and reconstructs 112. The reconstruction error is therefore -1.

For the next actual sample, 117, both sides predict the reconstructed value 112. The difference is 5, represented as 4, and the reconstructed value becomes 116. Again the error is -1.

If an agreed adaptation rule now changes the step to 2, an actual sample of 122 produces difference 6 and can reconstruct as 122. Both sides must make that step change together. This toy arithmetic illustrates synchronization and quantization; it is not G.726.

Separately, at 8,000 samples per second, four coded bits per sample give 32,000 payload bits per second. Packet headers and other transport overhead increase the network bandwidth beyond that figure.

Illustrative case study

Illustrative case study - fictional

A fictional service company migrates recorded maintenance calls into a searchable archive. Some recordings are labeled only ADPCM. One decoder produces recognizable but distorted speech, so the team initially suspects poor microphones.

Inspection reveals that the files use different ADPCM variants and require different decoding settings. The team obtains the format specifications, tests known recordings and checks sample rate, channel count and packing before transcription.

It keeps the original files and records the conversion settings beside each decoded copy. Correct format handling must precede judging speech-recognition accuracy.

Common misconceptions and mistakes

ADPCM does not mean one universally interchangeable file format.
Fewer bits do not imply exact preservation of the original waveform.
The predictor must use information reproducible at the decoder.
Adaptive quantization is different from changing a connection's network bitrate.
Decoding to a larger PCM file does not restore details already lost during compression.

Business use

ADPCM matters when integrating voice equipment, historical recordings or embedded audio products. Procurement and migration specifications should name the exact codec and container, not merely the family.

Test representative speech, background noise and sudden changes in level. Include repeated encode-decode cycles if the workflow transcodes recordings. For an AI application, evaluate the final task as well as listening quality: small audible differences and transcription errors are different measures. Preserve sample-rate information and verify channel handling so format mistakes do not masquerade as model weaknesses.

Check your understanding

Why should a predictive encoder base its next prediction on reconstructed samples rather than silently using original samples unavailable to the receiver?

Answer: The decoder needs to reproduce the same prediction. Different prediction histories can make the reconstructed stream diverge even when the transmitted difference codes are received correctly.

Related terms

Audio codec
pulse-code modulation
differential coding
quantisation
G.726
bitrate
speech recognition

Learn next

Compare the toy example with direct sample quantization. Then inspect a documented ADPCM format's state initialization and block boundaries, identifying what a decoder must know before it can reconstruct the first sample.

Sources

Sources: ITU-T G.726 recommendation and linked specification: https://www.itu.int/rec/T-REC-G.726-199012-I/en ; IETF RFC 3551, section 4.5.4: https://www.rfc-editor.org/rfc/rfc3551 . Relevant encoder, decoder and transport descriptions were reviewed. The invented arithmetic above is explanatory and is not a conformance test.

Previous: Adaptive design (medicine) Next: Adaptive dynamic programming