e-speak

Mechanics of Acoustic Echo Cancellation for Systems

By Nadia Ferrante

Most software documentation treats Acoustic Echo Cancellation (AEC) as a signal-processing black box containing adaptive finite impulse response filters and normalized least mean squares math. For an engineer operating at the system, driver, or protocol layer, that abstraction is useless.

AEC in production real-time communication is fundamentally a state machine and a buffer alignment problem. If your system passes corrupted reference audio, fails to account for ring-buffer latency in the operating system's audio server, or drifts across two asynchronous hardware clocks, the linear algebra inside the DSP engine fails immediately. The DSP does not fail because the acoustic model is wrong; it fails because the inputs violate the invariant assumptions of the cancellation loop.

This guide outlines how an echo canceller functions from an architectural standpoint, what states it transitions through, and where the host system breaks its assumptions.

What does the Reference Signal and the Problem of Alignment actually mean?

An echo canceller has two inputs and one output. The first input is the capture signal: raw pulse-code modulation (PCM) frames from the local microphone, containing local speech, background noise, and the sound emitted by the local loudspeaker. The second input is the reference signal: the PCM frames that were sent to the local digital-to-analog converter (DAC) to be played out through that loudspeaker. The output is the error signal: the capture signal minus the estimated echo, which is subsequently passed to the audio encoder.

` Far-End RTP ---> [Jitter Buffer] ---> [DAC / Speaker] --(Acoustic Path)--\ | | (Reference) v | [Microphone] v | [Echo Cancellation Engine] <------------------------- [ADC / Capture] | (Clean Audio) | v [Opus Encoder] ---> Near-End RTP `

The fundamental requirement of any cancellation engine is causality: an echo cannot be recorded by the microphone before the corresponding audio sample has physically left the speaker. If your audio pipeline delivers the reference frame to the canceller *after* the microphone frame containing that echo has already been ingested, cancellation is mathematically impossible.

The bulk delay window

In an operating system, audio render and audio capture run on distinct threads, often serviced by separate hardware buffers. When a 20 ms audio frame arrives over the network via RTP as specified in RFC 3550, it is decoded and pushed into a render buffer.

Between the moment those decoded samples are copied into the render ring buffer and the moment the pressure wave physically exits the speaker cone, hundreds of milliseconds can elapse. Operating system mixers, Bluetooth link-layer buffers, and hardware digital-to-analog converters introduce latency. The microphone pipeline introduces symmetric delays: analog-to-digital conversion, hardware DMA transfers, OS driver scheduling, and capture ring buffers.

The sum of the playback delay, the acoustic propagation through the air (roughly 3 ms per meter), and the capture delay is the bulk delay.

A standard echo canceller maintains an internal search buffer—frequently sized around 250 ms to 500 ms—to locate the reference audio. The cancellation engine correlates the capture stream against historical reference frames to determine the exact offset. If the bulk delay exceeds the depth of this buffer, the canceller cannot find the reference. It assumes there is no echo, leaves the microphone signal unattenuated, and the far-end participant hears their own voice returned to them unaltered.

Clock drift and buffer jitter

The reference signal and capture signal are rarely driven by the same physical clock crystal unless the endpoint is using a unified audio interface. If the loudspeaker is driven by a USB audio interface running at an actual rate of 48001 Hz, and the microphone is on an integrated motherboard codec running at 47998 Hz, the system exhibits clock skew.

A drift of 3 samples per second seems negligible to an application engineer, but to an echo canceller, it is catastrophic. A shift of even 1 ms (48 samples at 48 kHz) shifts the phase of the impulse response. The echo canceller's model of the room becomes misaligned, causing the linear filter to subtract energy from the wrong phase, which often amplifies the echo rather than canceling it.

When operating across two hardware clocks, the system audio engine must execute sample-rate drift compensation before feeding the reference buffer. This requires inserting or dropping samples—or applying polyphase sinc interpolation—to match the capture clock's ingestion rate. If the audio graph simply allows the reference queue to underrun or overrun, the step change destroys the filter state.

What does the State Machine of Echo Suppression actually mean?

An AEC does not apply a static subtraction formula. It runs an adaptive model controlled by a state machine that switches behavior based on the energy ratios of its inputs. The pipeline operates on fixed blocks, commonly 10 ms (480 samples at 48 kHz) or 20 ms frames.

Far-end single talk and the linear filter

When the far-end party is speaking and the near-end party is silent, the system is in Far-End Single Talk. The capture input consists exclusively of acoustic echo, room reverberation, and local ambient noise.

In this state, the adaptive linear filter does the primary work. It models the transfer function of the room—how sound bounces off walls, desks, and monitors—and generates an inverted synthetic echo. This synthetic signal is subtracted from the microphone signal. The efficiency of this subtraction is measured as Echo Return Loss Enhancement (ERLE). In a well-behaved room with a stationary microphone, a linear filter typically achieves 15 dB to 25 dB of attenuation.

Crucially, the linear filter *only* updates its coefficients during Far-End Single Talk. It requires a strong reference signal and an echo that matches that reference.

Double-talk detection and filter freeze

When both the far-end speaker and the local participant talk simultaneously, the system enters Double Talk. This is the hardest state for any echo canceller to handle.

The microphone now contains the acoustic echo plus the local voice. If the linear filter attempts to adapt its coefficients during double talk, it treats the local participant's voice as an error in its room model. The adaptation diverges within roughly 40 ms to 80 ms, completely corrupting the filter coefficients.

To prevent this divergence, the system employs a Double-Talk Detector (DTD). The DTD monitors the cross-correlation and power ratios between the reference signal, the capture signal, and the error signal. When it identifies energy in the capture channel that cannot be explained by the reference signal plus the known room response, the state machine fires an interrupt: it freezes filter adaptation instantly.

While the filter is frozen, it continues to subtract its existing model of the echo, but it does not update its coefficients until the near-end speaker stops talking.

Non-linear processing and comfort noise

Linear filters cannot eliminate 100% of an echo. Non-linearities in cheap speaker cones, physical vibration of the device chassis, room reverberation tails that exceed the filter length, and DAC/ADC harmonic distortion produce residual echo.

To eliminate this remaining 10 dB to 20 dB of sound, the AEC routes the subtracted signal through a Non-Linear Processor (NLP). The NLP is essentially an aggressive dynamic spectral suppressor. During far-end single talk, the NLP suppresses the residual output to near absolute silence.

However, completely zeroing the output causes the line to sound dead to the far-end participant, leading them to believe the call was dropped. To fix this, the AEC calculates the spectral profile of the local ambient room noise and injects synthetic comfort noise into the stream at an equivalent power level (often around -50 dBFS to -60 dBFS).

During double talk, the NLP must back off aggressively. If it clamps down while both parties speak, it chops off the words of the near-end speaker, causing half-duplex behavior reminiscent of legacy speakerphones. Tuning an AEC is almost entirely an exercise in balancing NLP aggression against near-end speech clipping.

Network Layer Failures and Transport Boundary Effects

Protocol engineering issues often manifest as phantom DSP bugs. The echo cancellation engine relies entirely on accurate time reporting from the network and playback pipeline.

"The timestamp reflects the sampling instant of the first octet in the RTP data packet."

— IETF RFC 3550, Section 5.1

If an application handles jitter buffering poorly, that guarantee breaks. When network jitter forces an RTP jitter buffer to suddenly drop a 20 ms frame or conceal a lost packet using interpolation as described in the Opus codec specification (RFC 6716), the rendered audio timeline changes. If the jitter buffer plays out a concealed frame, that concealed frame *must* be reflected identically in the AEC reference tap.

When packet loss exceeds 5% or network jitter spikes above 60 ms, poorly engineered media pipelines often drop or insert audio packets on the playback thread without notifying the capture pipeline. If the playback engine discards a single 10 ms audio buffer to catch up to the clock, but does not adjust the reference feed going to the AEC by the exact same 480 samples, the reference buffer and the physical room playback lose alignment. The bulk delay calculation instantly becomes invalid. The linear filter attempts to cancel current microphone input using past reference data, causing the filter to diverge, the double-talk detector to lock up, and loud echo bursts to leak back into the outgoing RTP stream for several seconds until the delay estimator can realign its cross-correlation window.

A related failure occurs at the hardware boundary with dynamic device switching. If an operating system shifts render output from a 48 kHz monitor speaker to a 16 kHz Bluetooth headset profile (HFP/HSP), the physical latency jump can exceed 150 ms in a single frame boundary. Unless the audio pipeline flushes the AEC reference queue, resets the filter taps, and forces the delay estimator back into an acquisition state, the echo canceller will attempt to adapt across two completely disparate acoustic transfer functions simultaneously.

Driver-level latency reporting is notoriously unreliable, particularly on desktop operating systems where WASAPI or ALSA drivers report theoretical hardware buffer sizes rather than real-time consumption pointers. I would check the specific audio backend implementation before assuming hardware presentation timestamps are true; in many cases, manually tracking the byte counters written to the sound card DMA buffer is the only way to avoid 30 ms to 50 ms of unmeasured jitter in your reference path.

The practical rule for systems engineers is straightforward: never attempt to fix echo leaks by tweaking filter algorithms or NLP gain curves until you have mathematically proven that the reference audio entering the canceller matches the physical output of the speaker down to the exact sample, with constant latency and zero uncompensated clock drift.

Terms used in this note

A codec is an algorithm that compresses audio for transmission. A jitter buffer is a queue that absorbs variable arrival times before playback. Packet loss concealment is the substitution a decoder performs when a frame never arrives.