What is AMR (Adaptive Multi-Rate)?

How AMR Works

AMR, short for Adaptive Multi-Rate, is a lossy speech codec built specifically to compress the human voice for mobile telephony rather than to reproduce general audio. Instead of treating sound as an arbitrary waveform, it models the voice using Algebraic Code-Excited Linear Prediction (ACELP), a technique in which the encoder estimates a mathematical model of the vocal tract and transmits only the parameters needed to drive that model rather than the raw samples themselves. This is why AMR can shrink speech dramatically while remaining intelligible: it is reconstructing a voice from a compact description rather than storing the sound.

The codec operates on short frames, each representing a small slice of speech, and for every frame it derives linear predictive coding (LPC) coefficients that describe the spectral shape of the voice, along with parameters describing pitch and an excitation signal chosen from a fixed algebraic codebook. Because speech has strong short-term and long-term correlation, this predictive approach removes a great deal of redundancy before any bits are transmitted, which is the core reason a voice call can sound acceptable at only a few kilobits per second.

The defining feature is in the name: AMR is adaptive and multi-rate. The narrowband version defines a set of bit rate modes ranging from very low rates up to around 12.2 kbps, and the encoder can switch between them on a frame-by-frame basis. When the radio channel is clean, it spends more bits on speech quality; when the channel degrades, it shifts bits away from the voice and toward stronger error protection, trading fidelity for robustness so the call does not break up.

AMR also includes tools tuned for real telephony. Voice Activity Detection (VAD) identifies silent gaps, and Discontinuous Transmission (DTX) stops sending full frames during those gaps, transmitting only occasional comfort noise parameters so the line does not sound dead. This saves battery and network capacity during the natural pauses of conversation. A wideband sibling, AMR-WB, extends the same design to a higher audio bandwidth for clearer, more natural sounding voice while keeping the same adaptive philosophy.

History and Standardization

AMR was developed to serve as a standardized speech codec for mobile networks and was adopted through the 3GPP, the body responsible for GSM and later 3G mobile standards. It emerged around the transition from second generation to third generation cellular systems, when carriers needed a single voice codec that could adapt gracefully to the variable, error-prone conditions of a wireless link. Making it a common standard meant handsets and network equipment from many vendors could interoperate on voice calls.

Because it was chosen as the mandatory or default voice codec across a generation of mobile networks, AMR became extraordinarily widespread on early mobile phones. Beyond live calls, it was used for on-device voice recording and for voice attachments in MMS messaging, which is why files with the .amr extension are still commonly encountered as recordings exported from older phones.

As mobile telephony moved toward all-IP networks and richer voice services, newer codecs took over the frontier, but AMR and AMR-WB remained deeply embedded in the ecosystem for a long time thanks to their entrenched hardware and network support. Their legacy is still visible today wherever compact voice files or backward-compatible telephony are involved.

AMR Compared to Music Codecs and Modern Speech Codecs

Compared with general-purpose music codecs such as MP3 or AAC, AMR is not a competitor so much as a different tool entirely. For speech at very low bit rates it is far more efficient and intelligible than those codecs, because its whole design assumes a single human voice. For music, however, it is essentially unusable: the vocal-tract model that makes speech so compact badly distorts instruments, harmony, and anything that is not a talking voice, so no one would choose AMR to store a song.

Against other speech codecs such as Speex, AMR occupied similar territory as a low-bitrate voice codec, though AMR had the advantage of being the built-in standard on mobile hardware while Speex saw more use in open-source and internet voice applications. The later arrival of Opus, a highly flexible codec that handles both speech and music across a wide range of bit rates, has largely superseded these older dedicated speech codecs for new internet and application voice work, offering better quality and a single codec for many use cases.

The honest tradeoff with AMR is that it is a narrow specialist that was right for its era and job. It delivers usable voice over unreliable, bandwidth-starved cellular links at a time when that was a hard problem, and it does so with modest processing requirements. Its weaknesses, poor music quality, limited audio bandwidth in the narrowband form, and its status as a legacy format with waning relevance for new work, all stem from the same narrow focus that made it excellent at the one thing it was built to do.