System Definition & Boundary (What this page covers)

Definition: A conference speakerphone combines a mic array, playback speakers, and a local DSP/SoC to deliver full-duplex voice over USB/BT/Ethernet. The critical engineering task is controlling the echo path and preserving speech clarity under realistic acoustic and electrical interference.

What “good” looks like (metrics you can verify)

• Far-end intelligibility (primary): stable clarity at distance and off-axis talker angles; avoid “thin/robotic” artifacts from overly aggressive NR/AGC.
• Echo suppression stability: high ERLE across voice band without collapsing during loud playback, EQ changes, or double-talk (barge-in).
• Latency budget discipline: total end-to-end delay kept within a range where AEC can converge; avoid hidden buffers and sample-rate conversions.
• Double-talk resilience: near-end speech remains present when far-end speaks; no “near-end gets canceled as echo.”

How this page is structured (evidence chain, not opinions)

• Signal flow: mic capture → beamforming → AEC/NR/AGC → transmit; and receive → playback → speaker.
• Reference integrity: where playback reference is tapped (pre/post EQ/limiter) and how delay is tracked.
• Power & coexistence: PoE/DC-DC/class-D switching noise vs mic AFE quiet domain.
• Validation mindset: measurable pass/fail checks (ERLE under stress, latency inventory, noise coupling correlation).

Acoustic Mechanics First (Enclosure, mic placement, speaker coupling)

Why this comes before algorithms: In real conference rooms, AEC and beamforming are limited by the echo path created by enclosure geometry, speaker-to-mic leakage, and tabletop reflections. When leakage dominates, “DSP tuning” often produces unstable, room-dependent behavior instead of repeatable gains.

Three coupling paths that set the AEC difficulty

• Airborne direct leakage: speaker energy escapes through ports/grilles and reaches the mic array with short delay—this raises residual echo and forces heavier suppression.
• Structure-borne vibration: enclosure or PCB vibration modulates mic signals (especially with tight mounts), producing echo-like components AEC cannot perfectly model.
• Tabletop early reflections: near-field reflections add strong early peaks in impulse response; they can confuse beamforming steering and reduce ERLE consistency across angles.

Mic array geometry: decisions with measurable consequences

• 2 mics: simpler, lower cost; limited spatial selectivity; more sensitive to reflections and talker movement.
• 4 mics (common sweet spot): stronger directivity and better noise rejection with moderate calibration needs.
• 6 mics: best spatial selectivity potential, but demands tighter channel matching (gain/phase), better clocking, and more robust mechanical symmetry.
• Spacing rule of thumb: larger spacing improves low-frequency directivity but increases spatial aliasing risk at higher frequencies; smaller spacing reduces aliasing but limits directivity.

Minimal evidence chain (fast, repeatable, hardware-first)

Discriminator: mechanics vs reference/latency

• Mechanics-dominant signature: AEC works in one room/position but fails dramatically after small placement changes (device rotation, table material, distance).
• Reference/latency signature: Echo behavior is consistently “wrong” across rooms, often tied to mode changes (USB vs BT) or audio pipeline changes (EQ/limiter order).

Mic Front-End (PDM/Analog mics, AFE noise, dynamic range, biasing)

Design goal: the mic front-end must preserve far-field speech while surviving loudspeaker leakage without saturating. Beamforming and AEC can only perform as well as the multi-channel input SNR and channel consistency allow.

PDM vs analog: choose based on the failure mode you can tolerate

What actually breaks far-field beamforming

• Self-noise + AFE noise floor: sets the intelligibility limit at distance. A clean DSP cannot recover information below the sensor/AFE noise floor.
• Sensitivity tolerance & phase mismatch: small gain/phase errors inflate sidelobes and blur the main lobe, reducing directional SNR gains.
• Leakage headroom: speakerphone mics must tolerate strong playback leakage; repeated clipping produces echo-like artifacts that confuse AEC and post-filters.
• PDM clocking artifacts: correlated clock/ground noise can appear in all channels at once, making “noise reduction” less effective because the noise is coherent.

AFE/ADC (or PDM receiver): three front-end requirements that matter in practice

• Low input-referred noise: keep equivalent acoustic noise low enough to resolve far-field speech in quiet and moderate-noise rooms.
• Headroom under leakage: avoid saturating on loud playback, coughs, table taps, or sudden near-field talkers; recovery behavior matters as much as peak level.
• Anti-alias & phase consistency: consistent filtering and group delay across channels protect beamforming performance; mismatched phase response reduces directional gain.

Matching & calibration hooks (designing for repeatability)

• Per-channel trim capability: reserve gain/phase trim or calibration tables in DSP for channel alignment (within reasonable correction range).
• Temperature drift awareness: bias networks and references drift; multi-channel mismatch can grow with temperature and aging, especially if layout is asymmetric.
• Symmetry by construction: identical trace lengths (where meaningful), consistent bias/RC networks, and consistent decoupling keep calibration stable over time.

Minimal measurement set (fast, repeatable, diagnostic)

Beamforming Pipeline (from raw channels to a clean talker)

Pipeline view: beamforming is a controlled sequence—align channels, apply delays/weights, sum, then suppress residual noise. Each stage has measurable outputs, and each stage fails in recognizable ways under mismatch, reflections, or near-field interference.

Fixed vs adaptive beamforming (select by constraints, not hype)

• Fixed beamforming: predictable latency and stable behavior; best when compute budget is tight and room conditions are moderately controlled.
• Adaptive beamforming: can track changing noise sources, but requires careful update control; mis-tuned updates can “chase reflections” and degrade speech naturalness.
• Speakerphone constraints: limited mic count, strict latency budget (AEC compatibility), and frequent early reflections from tables and walls.

Why VAD gating can destabilize beamforming

• Over-sensitive gating: speech tails get chopped; output sounds thin and intermittent, especially with soft talkers or off-axis speech.
• Under-sensitive gating: noise and leakage influence weight updates, creating steering drift or widened sidelobes.
• Practical guardrail: ensure gating decisions are consistent with AEC/double-talk logic; otherwise the system oscillates between “suppress” and “recover”.

Steering errors: three root causes that look similar but test differently

Evidence chain: 3 quick tests that isolate the limiting factor

• Angle sweep (polar sanity): fixed distance, rotate talker angle in steps; confirm a stable main lobe and acceptable sidelobes.
• Distance sweep (SNR vs distance): 0.5 m → 1 m → 2 m; if improvements collapse at distance, noise floor dominates (return to H2-3).
• Room/table A/B test: same setup with soft pad or different table; large improvement indicates reflection coupling dominates (return to H2-2).

AEC (Acoustic Echo Cancellation) Done Right

Core rule: AEC performance is limited less by “filter math” and more by whether the playback reference matches the echo component inside the mic signal in both content and timing. When reference is tapped at the wrong point or delay drifts, ERLE becomes unstable and barge-in fails.

Reference must be clean: where to tap (pre/post EQ, pre/post limiter)

Latency budget: hidden buffers and SRC are common ERLE killers

• Buffering: FIFO depth changes, mode switches (USB ↔ BT ↔ Ethernet), and safety buffers silently add delay.
• SRC (sample-rate conversion): introduces additional group delay and may drift under clock mismatch if not locked properly.
• Transport delays: USB frame scheduling, Bluetooth codec packetization, and network buffering can shift reference timing.
• Practical implication: AEC requires a stable and trackable delay between the reference and the mic echo; “mostly stable” often fails under stress.

Double-talk and residual echo: protect barge-in without destabilizing adaptation

• Double-talk detection: prevents the adaptive filter from updating aggressively while near-end and far-end speak simultaneously.
• Failure mode A (too strict): near-end speech is mistaken as echo → near-end gets suppressed, barge-in sounds weak.
• Failure mode B (too loose): filter updates during double-talk → divergence → ERLE collapses even after double-talk ends.
• Residual echo suppressor: cleans leftover echo after the main AEC, but excessive suppression trades intelligibility for “quiet”.

Nonlinear echo is real (speaker/amp/clipping): design for it

When nonlinear echo dominates, stable full-duplex requires both reducing nonlinearity (avoid clipping, manage limiter behavior) and post-AEC cleanup (residual suppression tuned to minimize speech damage).

Evidence: what to measure and which stress cases expose root causes

Noise Control Stack (NR/ANC/AGC, wind/fan, keyboard, room)

Goal: reduce background noise while preserving speech naturalness. Noise suppression is always a trade: higher SNR often increases artifacts unless processing order and dynamics are controlled.

Clarify terms: NR vs “ANC” in speakerphone context

• NR/NS (noise reduction/suppression): time-frequency suppression and post-filters applied to the speech signal.
• ANC: in strict terms requires secondary paths and error sensing; in many speakerphone discussions it loosely refers to suppression. This chapter focuses on noise suppression behavior.

Noise types drive different failure modes

AGC placement: mis-ordering can break AEC and destabilize the stack

• Dynamic blocks change the signal statistics: AGC and limiter reshape levels; if placed before AEC or in the wrong loop, they can degrade reference consistency.
• Common pitfall: AGC behavior that tracks far-end playback can interfere with AEC’s adaptation and double-talk logic.
• Practical rule: keep AEC’s view of reference and mic echo as stable as possible; apply aggressive dynamics after echo stability is secured.

Common artifacts and how to tie them to measurable indicators

Evidence: report SNR gains together with artifact cost

• SNR improvement vs distance: validate the gain at 0.5 m, 1 m, and 2 m; if improvements vanish at distance, noise floor dominates upstream (return to mic front-end).
• Artifact-cost metrics: track level variance, tonal residue, and time-varying spectral holes alongside subjective listening.
• Decision framing: prefer configurations that maintain intelligibility and stable full-duplex rather than maximizing “quietness”.

Playback Chain (DAC / Class-D amps / speakers / protection)

Design goal: deliver stable loudness and clean audio while minimizing leakage-driven artifacts that raise the workload for echo control. The playback chain must be treated as a controlled actuator: output power, distortion, noise, EMI, and thermal behavior all feed back into conferencing quality.

Class-D selection: choose by “bad-day behavior”, not by headline specs

• Output power vs supply: verify achievable power at the worst-case supply and speaker impedance; headroom matters more than peak numbers.
• THD+N vs power curve: near-max power distortion growth often dominates perceived harshness and echo residue.
• Idle noise (hiss): low-level noise becomes obvious in quiet rooms; check noise with playback muted and with system activity (USB/Ethernet/LED PWM).
• Spread-spectrum options: helps reduce EMI peaks; effectiveness depends on layout, filter components, and switching strategy.
• EMI reality: class-D edges can couple into mic/AFE grounds and RF paths; selection must consider switching frequency and control mode stability.

Speaker protection: the protection strategy is part of sound quality

Feedback and monitoring: make playback a closed-loop system

• Clip detect: triggers smooth gain reduction or EQ adjustment before hard clipping creates harsh artifacts.
• Temp sense: enables predictive thermal control (slow ramp-down) instead of sudden foldback.
• Current sense: detects abnormal loads and protects against overcurrent while providing a diagnostic signal for “bad cables/speaker faults”.
• DSP control: the cleanest solutions shape behavior upstream (avoid clipping) rather than relying on heavy post-processing.

Evidence: four minimal tests that explain most field complaints

Connectivity: USB Audio + Bluetooth Audio (and control plane)

System view: USB and Bluetooth are two I/O modes feeding the same DSP chain. The hardest problems are not “can it connect”, but clock-domain alignment, predictable latency, and clean mode switching without glitches, dropouts, or echo-control collapse.

USB Audio (UAC): hardware-focused failure modes

• Sample-rate handling: switching rates often triggers SRC and buffer resets; this can change end-to-end latency and break reference alignment.
• Clock domain crossing: USB-side timing and audio-processing timing must be bridged with stable buffering; under/overrun causes crackles or periodic dropouts.
• Frame/buffer effects: USB scheduling and internal FIFOs add fixed and sometimes variable delays; measure rather than assume.

Bluetooth: HFP vs A2DP and how codec latency impacts echo alignment

Control plane (mute buttons, LEDs, touch): common noise injection paths

• LED PWM: can modulate ground and supply rails, producing audible tones or raising the noise floor.
• Touch scanning: periodic excitation and long traces can couple into audio references if return paths are not controlled.
• Mute and mode keys: switching events must be sequenced to prevent pops/clicks and avoid transient reference disruption.

Evidence: measure mode switching and latency like a validation engineer

Ethernet + PoE: Power + Data Without Instability

Design goal: deliver reliable link-up and stable audio while Ethernet activity and PoE power conversion coexist in the same enclosure. Most “random reboot” and “Ethernet makes audio noisy” complaints reduce to inrush + PD/MPS behavior, start-up sequencing, and noise coupling paths between hot power zones and the mic/AFE domain.

PoE PD controller: four behaviors that decide field success

Isolation boundaries: keep “hot entry” and “audio quiet” physically and electrically separated

• Magnetics boundary: treat RJ45 + magnetics as a high-energy entry region with its own return paths.
• PoE front-end boundary: PD + primary conversion are a switching-noise source; keep them out of the mic/AFE reference area.
• Optional digital isolation (concept): used when the system must break ground loops or keep noisy domains from polluting sensitive references.

Noise coupling: the three most common paths into the mic domain

Evidence: three minimal tests that separate “power” from “coupling”

Power Tree & Grounding for Mixed-Signal Audio (the silent killer)

Core rule: most hiss, touch-induced noise, and “reboot at loud volume” issues are power/return-path problems. A mixed-signal speakerphone must treat the mic/AFE as a protected island and control where high di/dt currents return, especially from PoE conversion and class-D switching.

Rail partitioning: build islands and keep noisy domains out

• AFE rails: highest sensitivity; minimize ripple and prevent shared return paths with switching domains.
• Digital core: bursty current; isolate via local regulation and controlled return routes.
• RF domain: can both suffer from and inject noise; keep supply impedance stable.
• Class-D power: high di/dt; keep loops tight and returns away from AFE references.
• PoE/DC-DC: switching hot zone; prevent ripple and burst-mode noise from reaching analog rails.

Ground strategy: return-path control beats “pretty ground splits”

Transients: link-up, hot-plug, and bursts can trigger brownout loops

• Inrush & hot-plug: sudden load changes can sag the main rail and force repeated resets if UVLO thresholds are too tight.
• PoE burst behavior: light-load burst modes can create audible tones if ripple reaches sensitive rails.
• Peak load events: loud playback and compute spikes can expose insufficient margin and poor sequencing.

First 2 measurements: the fastest way to classify most failures

Discriminator patterns: correlate noise with the real aggressor

Fix priority: isolate first, then stabilize events, then polish features

• Isolation: protect the AFE island and keep noisy returns out.
• Event stability: tame inrush/hot-plug/link-up and choose UVLO/brownout behavior with margin.
• Polish: minimize LED/touch injection and interface coexistence issues once fundamentals are clean.

EMC/ESD/Audio Coexistence (pass tests and sound good)

Goal: survive ESD/EFT and radiated/conducted stress without muting, rebooting, or degrading near-end voice. Practical success comes from mapping aggressors → coupling → victims, then proving fixes with correlation (near-field scans, hit maps, and audio artifact logs tied to test points).

Typical aggressors (where the energy starts)

• Class-D outputs: high dv/dt and speaker cable radiation; edges can capacitively inject into mic inputs.
• PoE DC/DC edges: switching ripple and burst-mode behavior can become audible if rails leak into AFE references.
• Ethernet PHY activity: traffic bursts modulate digital return currents; noise can track activity level.
• USB cable entry + ESD clamps: the clamp current return path can create ground bounce that upsets refs, clocks, or reset lines.

Typical victims (where small signals break first)

• Mic inputs: high impedance, low signal level; susceptible to E-field coupling and common-impedance return noise.
• AFE references / bias nodes: ripple or bounce here turns into broadband hiss or tonal artifacts.
• Clocks: disturbance can cause SRC/DSP timing artifacts, dropouts, or mode-switch instability.
• Touch / LED lines: scanning/PWM can both inject noise and latch-up/reset under ESD if entry protection is weak.

Layout tactics that preserve audio quality (hardware-first)

Evidence: make interference visible and repeatable

Concrete example MPNs (typical “first picks”)

These are commonly used reference parts to speed selection. Final choice depends on voltage, IEC level targets, capacitance limits, and layout constraints.

Validation & Field Debug Playbook (symptom → evidence → isolate → fix)

Format: each symptom is handled with the same repeatable sequence: First 2 measurements (minimum tools) → Discriminator (what proves the root-cause category) → First fix (highest leverage change). Component-level examples are included to accelerate troubleshooting and redesign loops.

1) Far-end hears echo / ERLE low

• First 2 measurements: (A) capture the playback reference at the chosen tap point (pre/post EQ/limiter), (B) measure end-to-end delay change between modes (USB vs BT) using an impulse.
• Discriminator: ERLE collapses only at high volume → nonlinear playback (clipping/limiter) contaminates the echo; ERLE shifts after mode switching → delay/buffer reset misalignment.
• First fix: move reference tap to a stable point; align delay after SRC/buffer; avoid hard limiting in the reference path.
• Example MPNs (playback/telemetry): TI TAS5825M, TI TAS5805M, Maxim MAX98390 (choose based on EMI/noise/protection telemetry needs).

2) Near-end voice thin/robotic (NR artifacts)

• First 2 measurements: (A) compare spectra pre/post NR+AGC, (B) log output level variance vs time (pumping) during steady speech/noise.
• Discriminator: artifacts increase with fan/keyboard noise → aggressive suppression; artifacts appear after CPU load or streaming changes → timing/buffer stress or order issues.
• First fix: enforce processing order stability (avoid AGC breaking AEC); relax thresholds/time constants; cap suppression aggressiveness before chasing “more DSP”.
• Hardware helpers (noise immunity): low-noise AFE rail regulation often reduces “NR overreaction” by lowering the raw noise floor: TI TPS7A20, ADI LT3042.

3) Howling at high volume (acoustic coupling + limiter + AEC instability)

• First 2 measurements: (A) record mic input and speaker output simultaneously to locate feedback frequency, (B) monitor clip/temp/current flags at the amp stage.
• Discriminator: fixed strong frequency → acoustic/mechanical resonance; howling appears at limiter action → nonlinearity plus AEC instability.
• First fix: reduce leakage (mechanical/placement) and stabilize limiter behavior (smooth control, avoid hard clipping); validate at worst-case volume.
• Example MPNs: amps with robust protection hooks: TI TAS5825M, Maxim MAX98390.

4) Touch/LED causes hiss or clicks

• First 2 measurements: (A) AFE rail ripple near the AFE load, (B) correlate ripple/noise with LED PWM or touch scan timing.
• Discriminator: noise matches PWM frequency/harmonics → conducted/return coupling; click at touch events → transient injection via return path or clamp current path.
• First fix: isolate returns; slow edges or move PWM frequency; add rail filtering close to AFE; ensure clamp currents return outside the quiet reference region.
• Example MPNs (rail isolation): load switch TI TPS22919 (domain gating), low-noise LDO TI TPS7A20.

5) PoE negotiation resets / reboots on link-up

• First 2 measurements: (A) PoE main rail / 5V sag during link-up and start-up, (B) reset/PG/UVLO behavior (is it a brownout loop?).
• Discriminator: repeated ramp-up then drop → inrush/MPS/power budget; only fails on certain switches → boundary/tolerance sensitivity (often inrush behavior).
• First fix: tune inrush and start sequencing; add hot-swap/eFuse if needed; set supervisor thresholds/hysteresis for real sag events.
• Example MPNs: PD controller TI TPS2372/TPS2373, hot-swap/eFuse TI TPS25942, supervisor TI TPS3839.

6) BT drops when speaker plays loud (ground bounce / RF detune)

• First 2 measurements: (A) main rail sag vs class-D current events (volume steps), (B) RF/BT supply noise or ground reference disturbance under loud playback.
• Discriminator: dropouts scale with volume more than distance → power/return bounce; dropouts change with hand position/enclosure → antenna detune sensitivity.
• First fix: keep class-D return away from RF domain; add transient headroom; stabilize RF rail; reduce edge coupling.
• Example MPNs (power stability): eFuse TI TPS25940, low-noise LDO TI TPS7A20 for sensitive rails (as applicable).