Modern True Wireless Stereo (TWS) SoCs: A Technical Exposition

 

1. System Definition and the Constraint Envelope

1.1 What a TWS System Actually Is

A TWS product is not a headphone with a cut cable. It is a distributed, battery-powered, two-node wireless audio system with a third, quasi-dormant node (the case), in which two physically separate transceivers must:
  • maintain independent or coordinated links to a source device;
  • render a sample-coherent binaural image despite having no shared clock domain by default;
  • execute closed-loop acoustic control (ANC) with microsecond-scale latency;
  • perform multi-microphone speech enhancement, sensor fusion, and increasingly neural inference;
  • do all of the above within an energy budget smaller than a smartwatch's and a volume smaller than a sugar cube.

1.2 The Constraint Envelope

Constraint
Typical magnitude
Architectural consequence
Battery energy per bud
40–80 mAh at ~3.7 V ≈ 150–300 mWh
Everything is duty-cycled; average system power must sit near 20–40 mW for 6–8 h runtime
Electronics volume
~1–3 cm³ total; PCB often <1 cm²
System-in-package (SiP), embedded flash, minimal passives, antenna co-design
Thermal/skin contact
Sustained dissipation perceptible above ~100 mW
No fanout of phone-class compute; mW-per-GOPS discipline
RF environment
Antenna millimeters from lossy head tissue; cross-head shadowing of tens of dB at 2.4 GHz
Link diversity, mirroring/snooping topologies, aggressive retransmission engineering
Acoustic geometry
Mic-to-eardrum distances of ~1–2 cm
Causality bounds on feedforward ANC (tens of microseconds)
BOM economics
SoC ASPs from <$1 (value tier) to several dollars (flagship)
Tiered portfolios; RISC-V and mature-node cost engineering at the low end
The central design thesis of every modern TWS SoC follows from this table: it is a mixed-signal, multi-domain, duty-cycle-optimized control-and-audio computer with a radio attached — not a scaled-down smartphone modem.

2. Canonical Architecture of a Modern TWS SoC

2.1 Functional Block Decomposition


The blocks merit individual treatment:
  1. Radio front end. A dual-mode (Bluetooth Classic BR/EDR + Bluetooth LE) 2.4 GHz transceiver with on-die PA/LNA, fractional-N synthesizer, and digitally assisted calibration. Leading platforms now add a second radio: Qualcomm's Snapdragon Sound Elite Gen 2 integrates micro-power Wi-Fi 6E across 2.4/5/6 GHz on the same chip.
  2. Link controller. Hardware/firmware state machines that execute piconet scheduling, ISO event framing, retransmission windows, and encryption with minimal CPU wakeups — the single largest lever on standby and connected-idle power.
  3. Host/application CPU. Typically Arm Cortex-M33 (Airoha AB1595 uses a Cortex-M33F host) or Cortex-M55 (Qualcomm Elite Gen 2 uses an M55 at 300 MHz as the application "housekeeper"; Bestechnic's BES2800BP integrates a dual-core Cortex-M55). TrustZone-M isolation and secure boot are now table stakes.
  4. Audio DSP cluster. Cadence Tensilica HiFi-class cores dominate: the AB1595 carries dual HiFi 5 cores plus Airoha's own "AiroHAP" audio hardware processor; Qualcomm's S7 Gen 1 fields 2×500 MHz + 1×250 MHz DSPs, and Elite Gen 2 doubles the HiFi resources (2× HiFi 5 at 500 MHz) while adding a Helium-capable M55 at 500 MHz as main audio DSP and a HiFi 3 at 250 MHz dedicated to low-latency work.
  5. Neural accelerator. Qualcomm's embedded NPU delivers 64 GOPS on S7 Gen 1 and 128 GOPS (256 MACs/cycle at 250 MHz) on Elite Gen 2 [[50]; Airoha markets the AB1595 as its first flagship with a built-in AI hardware accelerator.
  6. Mixed-signal audio. Multi-channel sigma-delta ADCs (high dynamic range for ANC/voice microphones), hi-fi stereo DACs — Qualcomm specifies a "Hi-Fi-grade stereo audio codec" supporting up to 24-bit/192 kHz — and cap-less earpiece drivers that eliminate bulky DC-blocking capacitors.
  7. PMU and charger. Integrated buck/boost conversion, multiple LDO domains, Li-ion CC/CV charging, BC1.2 detection (explicitly listed on Airoha parts), and power-line or wireless case communication for state-of-charge, pairing, and DFU coordination.
  8. Memory. The generational jump is stark: S5 Gen 2 (QCC5181) carries 2.64 MB RAM versus 10.6 MB on S7 Gen 1 — a deliberate ~3× increase to host neural models and larger audio graphs. Airoha embeds 8/16 MB flash on the AB1595.

2.2 Heterogeneous Compute: The Workload-Placement Doctrine

The most instructive public articulation of modern TWS compute partitioning comes from Qualcomm's Elite Gen 2 disclosure: very light AI runs on the Arm M-class core, moderate neural work on the HiFi DSPs, and heavy inference on the NPU, while deterministic classical audio (EQ, DRC, filters) stays on DSP, with an M-class "housekeeper" scheduling the framework. This is codified in Qualcomm's Kymera multi-core audio architecture, which first appears in the S7 Gen 1 block diagram and is productized as the "Kymera Multi-Core Audio Architecture" in Elite Gen 2.
Workload
Preferred engine
Rationale
ANC fixed filters, biquad cascades
Dedicated low-latency DSP/HW
Microsecond group delay; determinism
Codec encode/decode, EQ, DRC, ASRC
HiFi-class DSP
MAC-dense, well-vectorized, mature libraries
DNN noise suppression, KWS, scene classification
NPU
Best energy per MAC; frees DSP headroom
Adaptive-filter control, policy, mode arbitration
Cortex-M CPU
Branchy control code; occasional light ML (Helium)
Sensor fusion (IMU, in-ear, PPG front-end)
Sensor hub / always-on domain
Runs while application domain sleeps
The architectural consequence, per Qualcomm, is that the audio-and-AI subsystem can now fetch, prepare, and execute models alongside classical processing without bottlenecking the application processor — the defining difference between "DSP with AI features" (2023) and "AI-native audio computer" (2026).

2.3 Documented Exemplars (Verified Specifications)

Platform
Announced
Host CPU
Audio DSP
AI compute
Memory
Radios
Qualcomm S5 Gen 2 (QCC5181)
2022
n/d
2×240 MHz
— (DSP-based)
2.64 MB RAM
BT dual-mode, LE Audio
Qualcomm S7 / S7 Pro Gen 1 (QCC7226 / +QCP7321)
Oct 2023
App processor + sensor hub
2×500 MHz + 1×250 MHz
eNPU, 64 GOPS
10.6 MB RAM
BT 5.4 + LE Audio/Auracast; Pro adds companion micro-power Wi-Fi, up to 29 Mb/s XPAN
Qualcomm Snapdragon Sound Elite Gen 2
Sep 2026
Cortex-M55 @ 300 MHz
M55 @ 500 MHz + 2× HiFi 5 @ 500 MHz + HiFi 3 @ 250 MHz
eNPU, 128 GOPS
n/d
BT 7.0 incl. HDT (~7.5 Mb/s) + integrated tri-band Wi-Fi 6E
Airoha AB1595
Dec 2024
Cortex-M33F
Dual HiFi 5 + AiroHAP
AI HW accelerator
8/16 MB embedded flash
BR/EDR + LE Audio + proprietary "Bluetooth High Data Rate"
Bestechnic BES2800BP
BES2800 series 2023
Dual Cortex-M55
Optional DSP
NPU
n/d
BT 5.4 + Wi-Fi 6 combo
Apple H2 (AirPods Pro 3)
Sep 2025
Undisclosed
Undisclosed ("computational audio")
Undisclosed
Undisclosed
BT dual-mode + proprietary protocols; H2 confirmed in Apple specs
Two integration milestones deserve emphasis:
  • Process and area. Bestechnic announced the BES2800 series as a 6 nm FinFET ultra-low-power wearable SoC in 2023 — the leading edge of the category. Qualcomm reports Elite Gen 2 at a combined SoC area of 31.185 mm², ~30% smaller than the S7-generation two-chip Wi-Fi solution, achieved by absorbing the Wi-Fi companion (QCP7321) onto the main die.
  • From two chips to one. The S7 Pro Gen 1 was explicitly a QCC7226 + QCP7321 pair; Elite Gen 2's single-chip integration is what makes Wi-Fi-equipped earbuds and glasses mechanically and economically routine.

2.4 Market Strata

  1. Flagship merchant: Qualcomm S7/Elite; Airoha AB1585/AB1595 (AB1585: BT 5.3, LE Audio certified, HiFi 5 DSP).
  2. Mid-tier merchant: Qualcomm S3/S5 Gen 3 (March 2024; BT 5.4, LE Audio + Auracast across the family, third-party customization on S3 Gen 3); Airoha AB157x (Cortex-M4F, BT 5.3, LE Audio); Bestechnic BES2700-class.
  3. Value tier: Airoha AB1562 (BT 5.3 dual-mode, hybrid ANC), Realtek RTL8763/8773 families, and the high-volume Chinese vendors (Bluetrum, JieLi, Actions), increasingly on RISC-V cores at mature nodes — the segment that made ANC-equipped TWS a sub-$20 commodity.
  4. Captive/OEM: Apple H2, Huawei Kirin A2/A3, Google Tensor A1, Sony V2+QN — silicon co-designed with a proprietary device ecosystem (Sections 6–7, 10).

3. The Connectivity Subsystem

3.1 Dual-Mode Radio and Throughput Physics

Modern TWS SoCs remain dual-mode: Bluetooth Classic (BR/EDR) for A2DP/HFP interoperability with the installed base, and Bluetooth LE for LE Audio, control, and proximity features. Qualcomm is explicit that Elite Gen 2 remains dual-mode even with High Data Throughput added — HDT is an extension of LE, not a third mode.
Classic ACL capacity. With EDR, the canonical asymmetric maximum user throughputs follow directly from payload sizes over a six-slot (3.75 ms) forward/reverse cycle:
DH5 : 339 bytes × 8 / 3.75 ms = 723.2 kb/s 3-DH5 : 1021 bytes × 8 / 3.75 ms = 2178.1 kb/s
Thus ~2.1 Mb/s is the theoretical Classic ceiling; practical A2DP sustain rates are materially lower once retransmissions, L2CAP/AVDTP overhead, and Wi-Fi coexistence are accounted for. This is precisely why LDAC's 990 kb/s mode and aptX Lossless's ~1.2 Mb/s envelope are RF-fragile, and why adaptive codecs exist (Section 4).
Link budget. A first-order earbud link budget:
P_rx = P_tx + G_tx + G_rx − L_fs − L_body L_fs(2.44 GHz, 1 m) ≈ 40 dB; L_body (cross-head/torso) ≈ 20–35 dB
With typical realized antenna gains of roughly −10 to 0 dBi in-ear, +4 to +10 dBm transmit, and receiver sensitivities near −95 dBm (LE 1M class), the margin is adequate for pocket-to-ear links but thin for cross-head ear-to-ear links — the physical reason topologies below evolved away from bud-to-bud relaying.

3.2 The Topology Problem and Its Four Historical Solutions

The defining TWS-specific protocol problem: delivering two synchronized channels to two independent radios.
Topology
Era
Mechanism
Strengths
Weaknesses
Master–slave relay
~2016–2018
Phone links to master bud; master forwards the other channel over a second link
Works with any phone
Double airtime, asymmetric battery drain, added latency, cross-head link fragility
Snoop / mirroring
~2018–2022
Both buds observe one piconet; one ACKs; roles swap seamlessly (Qualcomm TrueWireless Mirroring, Airoha MCSync, Apple's proprietary scheme). Variant: TrueWireless Stereo Plus dual independent links (phone-side support required)
Single visible device; balanced drain; no relay airtime
Vendor-proprietary; role-swap and key-sharing complexity
LE Audio unicast (CIS/CIG)
2022–
Two Connected Isochronous Streams in one group, anchored to a common controller clock with a negotiated presentation delay; bounded retransmission via flush timeout
Standardized, interoperable, symmetric power, multi-device (CAP/CSIP)
Requires LE Audio on both ends; ecosystem transition
LE Audio broadcast (BIS/BIG, Auracast)
2023–
Connectionless isochronous broadcast; unlimited receivers; open or Broadcast_Code-encrypted
Public audio, assistive listening, shared listening
No per-receiver ACK (reliability via repetition); assistant UX needed
LE Audio's isochronous architecture is the first standards-native solution to the binaural problem: both earbuds rendezvous at a shared presentation timestamp rather than relying on vendor mirroring tricks. Android has supported LE Audio natively since Android 13, and the profile stack — BAP/CAP/TMAP/GMAP/HAP/PBP with CSIP for set identification — is now the default design target for new silicon.

3.3 Binaural Synchronization Engineering

Even with a shared anchor, each bud's audio clock drifts. The arithmetic is unforgiving:
20 ppm crystal error → 20 µs/s → 72 ms of drift per hour Human ITD acuity ≈ 10–20 µs → image shift long before drift is "large"
Consequently, TWS SoCs implement continuous sample-rate reconciliation (asynchronous sample-rate conversion or fractional sample insertion/removal) slaved to the Bluetooth anchor clock, plus timestamped DAC start alignment. Mirroring-era designs achieved the same via shared piconet clock observation and periodic resync handshakes; LE Audio formalizes it via presentation delay. This block — not the codec — is where binaural image stability is won or lost.

3.4 Beyond Bluetooth: The Multi-Radio Turn (2023–2026)

Four parallel extensions attack Bluetooth's ~2 Mb/s practical ceiling:
  1. Qualcomm XPAN / micro-power Wi-Fi. Introduced on S7 Pro Gen 1: up to 29 Mb/s, enabling 24-bit/192 kHz lossless over Wi-Fi at earbud-class power, with whole-home coverage and seamless Bluetooth↔Wi-Fi transitions. Elite Gen 2 integrates tri-band Wi-Fi 6E and targets direct-to-cloud streaming — earbuds with their own IP connectivity, no phone in the audio path — with a simulated Wi-Fi-connected idle target of 0.6 mA (down from a 1.0 mA comparison target; vendor simulation, pre-silicon). An ecosystem of on-device "apps and agents" (Cinemo, Qobuz, Roon; MicroEJ runtime) is being built atop it.
  2. Huawei NearLink + L2HC. The FreeBuds Pro 4 (Kirin A2) introduced L2HC 4.0 with up to 2.3 Mb/s lossless over upgraded NearLink; the FreeBuds Pro 5 (Kirin A3, November 2025) continues NearLink audio at 2.3 Mb/s with compatible Huawei phones.
  3. Proprietary high-rate LE extensions. Airoha documents an "AIROHA/MediaTek Bluetooth High Data Rate" mode on the AB1595; Qualcomm had long shipped QHS as a proprietary high-speed link.
  4. Standardization catches up: Bluetooth 7.0 / HDT. On November 4, 2026, the Bluetooth SIG adopted High Data Throughput, an optional Core 7.0 feature raising the LE maximum bit rate to 7.5 Mb/s — nearly 4× the LE 2M PHY — explicitly setting the stage for a standardized lossless audio profile. Qualcomm's Elite Gen 2 already announces BT 7.0/HDT support with fallback to LE 1M/2M under congestion. This is the pivotal event of the current cycle: the proprietary lossless arms race (aptX Lossless, L2HC, LHDC, QHS) is being folded into interoperable standard transport.

3.5 Standards Timeline (Reference)

Release
Date
TWS-relevant content
Core 5.2
Jan 2020
LE Isochronous Channels, EATT, LE Power Control — LE Audio foundation
LE Audio spec set
Completed 2022
LC3, BAP/CAP/TMAP/GMAP/HAP/PBP, CSIP
Core 5.3 / 5.4
2021 / 2023
Connection subrating-class efficiency; PAwR, encrypted advertising
Core 6.0
Sep 2024
Channel Sounding — secure fine ranging
Core 6.1
May 2025
Randomized RPA timing offload — privacy + controller-side power savings
Core 6.2
Nov 2025
Minimum connection interval to 375 µs (125 µs granularity); Channel Sounding attack resilience
Core 6.3
May 2026
CS Inline PCT Transfer; PHY-specific RTT accuracy; HCI mask expansion ("Running Out of Bits"); BR/EDR↔LE ACP/C/I limit alignment
Core 7.0
Nov 4, 2026
HDT: 7.5 Mb/s LE PHY; standardized lossless music profile pathway
Channel Sounding (6.0–6.3) is quietly strategic for TWS: centimeter-class secure ranging underpins precision case-finding and proximity-aware audio switching, functions previously requiring UWB companions.

4. Audio Codecs: The Rate–Robustness–Latency Trilemma

4.1 Codec Landscape

Codec
Core technique
Max resolution
Bitrate envelope
Transport
Notes
SBC
Subband ADPCM
48 kHz/16-bit
~328–345 kb/s (stereo)
A2DP (mandatory)
Universal fallback; modest quality per bit
AAC-LC
MDCT psychoacoustic
48 kHz/16-bit (typ.)
~256 kb/s
A2DP
Encoder-quality variance dominates outcomes
aptX / aptX HD
ADPCM
48 kHz/24-bit (HD)
~352 / 576 kb/s
A2DP
Deterministic latency; Qualcomm-licensed
aptX Adaptive
Hybrid adaptive
96 kHz/24-bit
~279–420 kb/s, dynamically scaled
A2DP
RF-aware rate + gaming mode
aptX Lossless
Adaptive + lossless mode
44.1 kHz/16-bit lossless; 96 kHz/24-bit lossy
up to ~1.2 Mb/s
A2DP (Snapdragon Sound)
Bit-exact CD when RF permits; designed to scale with link quality
LDAC
Hybrid MDCT
96 kHz/24-bit
330/660/990 kb/s
A2DP
Sony; 990k mode RF-fragile
LHDC 5.0
Hybrid
192 kHz/24-bit
up to ~1 Mb/s
A2DP
Savitech; strong in Chinese OEM ecosystem
LC3
MDCT + entropy coding
48 kHz/24-bit; 8–48 kHz rates
~16–192+ kb/s per channel
LE Audio (mandatory)
7.5/10 ms frames; surpasses SBC at roughly half the rate
LC3plus
LC3 superset (ETSI TS 103 634)
96 kHz/24-bit
flexible
Proprietary/DECT; licensed
2.5/5 ms frames; FEC for erasure channels
L2HC 4.0
Huawei proprietary
Lossless
up to 2.3 Mb/s
NearLink/proprietary
Requires Huawei source device
Samsung SSC (UHQ)
Scalable proprietary
24-bit class
adaptive
A2DP/proprietary
Galaxy ecosystem

4.2 Why LC3 Is the Inflection Point

LC3's design brief was economic, not audiophile: transparent-ish quality at half the SBC bitrate, freeing airtime for retransmissions and dual-stream ISO scheduling. Its MDCT front end, long-term/post-filtering tools, and arithmetic coding deliver at ~96 kb/s per channel what SBC needs ~200+ kb/s to approach; its 7.5/10 ms frame grid plus bounded flush windows make latency a negotiated parameter (presentation delay) rather than an emergent accident of jitter buffers. Combined with SWB voice (32 kHz) under TMAP — and LC3-based super-wideband in late HFP revisions — LC3 unifies music and voice coding on one engine, which is exactly what a constrained DSP/NPU budget wants.

4.3 The Lossless Question, Honestly Framed

  • 16-bit/44.1 kHz PCM = 1.411 Mb/s raw; entropy-reduced music typically compresses to ~0.7–1.1 Mb/s — right at the edge of Classic EDR's practical envelope, hence aptX Lossless's RF-aware toggling between bit-exact and high-rate lossy.
  • 24-bit/192 kHz = 9.2 Mb/s raw — beyond any Bluetooth Classic mode; this is why Qualcomm routed it over XPAN Wi-Fi (29 Mb/s), Huawei over NearLink (2.3 Mb/s compressed lossless at CD-class rates), and why HDT's 7.5 Mb/s matters: it makes hi-res lossless plausible over standardized Bluetooth for the first time.
  • The perceptual caveat belongs in any rigorous treatment: the transport can already exceed what the acoustic transducer and the ear resolve; Qualcomm's own product management concedes that rendering, not transport, is the binding constraint at these rates.

4.4 Latency Budgets

End-to-end latency decomposes as:
T_e2e = T_encode(frame + lookahead) + T_air(scheduling + retransmissions) + T_jitter/presentation buffer + T_decode + T_DAC/driver
Mode
Typical end-to-end
Engineering note
A2DP + SBC/AAC
~150–250 ms
Large jitter buffers absorb Classic's stochastic retransmission
aptX Adaptive gaming mode
~50–80 ms
Rate sacrificed for buffer depth
LE Audio unicast (music configs)
tens of ms, config-dependent
Presentation delay is negotiated, not accidental
LE Audio GMAP
≤40 ms total system delay (Level A requirement)
Mandatory codec configs + QoS make "low latency" a testable claim
Proprietary 2.4 GHz dongles
~20–30 ms
Dedicated scheduler, no coexistence with phone traffic
Wired reference
<5 ms
The perceptual gold standard for monitoring
(Ranges are engineering-typical, not specifications, except GMAP's normative bound.)

5. ANC and Transparency: Real-Time Acoustic Control on a Milliwatt Budget

5.1 Topologies

  • Feedforward (FF): external mic senses noise before it reaches the eardrum; an anti-noise filter drives the speaker. Broadband potential, but causal: the electronics must beat the acoustic propagation delay.
  • Feedback (FB): mic inside the cavity, near the driver, closes a loop on the residual error. Robust to fit variation at low frequencies; bandwidth-limited by loop delay.
  • Hybrid: FF + FB — the universal flagship configuration (explicitly supported in silicon, e.g., Airoha AB1595's FF+FB+real-time adaptive ANC).

5.2 The Causality Constraint (FF)

τ_ADC + τ_filter + τ_DAC + τ_driver ≤ τ_acoustic(mic → eardrum) ≈ d/c d ≈ 1–2 cm → τ_acoustic ≈ 30–60 µs
This is why ANC cannot be "just another DSP task": the entire conversion-and-filter chain must complete in tens of microseconds. Qualcomm's S7-generation answer is a dedicated, fully programmable 3 µs low-latency DSP path alongside 112 biquad filters in its 4th-generation adaptive ANC hardware — a concrete, published anchor for how tight these budgets are. Elite Gen 2 extends this to a 5th-generation, AI-assisted adaptive ANC.

5.3 Feedback Loop Limits and the Waterbed

The FB loop's secondary path (DAC→driver→cavity→mic→ADC) imposes delay τ_s; unity-gain crossover must sit well below the frequency where accumulated phase lag exhausts margin — practically ~1–2 kHz. Bode's sensitivity integral then guarantees the "waterbed": suppression below crossover is paid for by noise amplification above it, which competent designs keep ultrasonic or masked. This is the control-theoretic reason hybrid ANC exists and why "ANC depth" claims are meaningless without a bandwidth and a measurement standard (Apple, notably, benchmarks its AirPods Pro 3 "2× more noise removed" claim against IEC 60268-24).

5.4 Adaptation: From Fixed Filters to Fit-Aware Control

Real ears leak. Modern ANC therefore runs adaptive filtered-x LMS variants with leakage:
w(n+1) = (1 − µγ)·w(n) − µ·e(n)·x′(n), x′(n) = ŝ(n) ∗ x(n)
where ŝ is the estimated secondary path, µ the step size, γ the leakage factor stabilizing the loop under model mismatch. Shipping silicon documents exactly this class of behavior: the AB1595 detects wearing condition (good fit vs. leakage) and auto-adjusts internal filters per environment while balancing "ear pressure" comfort; Qualcomm's 4th-gen ANC adapts to fit loosening and environmental transitions, with AI-based scene/mode arbitration added at Elite Gen 2. Sony's flagship line goes further architecturally, pairing the V2 Bluetooth-audio processor with a separate QN-series ANC processor — the WF-1000XM6's QN3e is claimed to be 3× faster than the XM5's QN2e, with an Adaptive Noise Canceling Optimizer analyzing noise and wearing conditions in real time.

5.5 Transparency and Own-Voice

Transparency is ANC's dual problem: re-inject the world with near-zero group delay and flat magnitude so occlusion disappears and one's own voice sounds natural. It stresses the same low-latency path (hence HiFi 3-class dedicated cores) and is where per-unit factory calibration (mic sensitivity/phase, driver response) most audibly pays off. Apple's AirPods Pro 3 iteration illustrates the system-level co-design: a new multiport acoustic architecture and foam-infused tips raise passive isolation, letting computational ANC and personalized Transparency do less work for more effect.

6. Voice Uplink and On-Device AI

6.1 The Hardest Signal Path in the Product

Uplink speech must be extracted at the mouth-far end of a head-worn device, in wind and 80+ dB babble, from 2–4 tiny MEMS microphones per bud. The modern chain:
  1. Spatial filtering: fixed/differential beamforming across per-bud mics; binaural (cross-bud) fusion where link budget allows.
  2. Bone-conduction/VPU fusion: accelerometer-based voice pickup gates the array in extreme noise.
  3. DNN enhancement: neural suppression of non-stationary noise and residual echo — the workload that justified NPUs in this category.
  4. AEC and sidetone management, wind detection with structural (mesh/port) and algorithmic mitigation.
The published state of the art is aggressive: Airoha claims the AB1595, coordinating up to 10 microphones with AI voice separation, pushes speech noise suppression from ~10 dB (consumer baseline) to up to 40 dB, targeting Microsoft Teams Open Office-grade quality on boomless headsets. Qualcomm's account of AI cVc is architecturally instructive: a hybrid design where classical, deterministic, low-power algorithms remain, and neural components moderate and adapt them (e.g., filter selection per environment) — with some customers running fully black-box neural enhancement at very low latency.

6.2 Captive Silicon and the Compute Marketing Wars

  • Google Tensor A1 (Pixel Buds Pro 2, 2024): the first Tensor in an earbud, marketed as running "90 times faster than the speed of sound" — a metaphor for its dedicated audio pipeline — powering a claimed 2× ANC improvement; Google brought it down-market to the Pixel Buds 2a in 2025.
  • Apple H2 (AirPods Pro 2 → AirPods Pro 3, 2025): Apple's "computational audio" framing; on AirPods Pro 3 the H2's pipeline underpins 2× ANC, next-gen Adaptive EQ, and Live Translation rendering (translation inference itself runs on the paired iPhone's Apple Intelligence).
  • Qualcomm eNPU: 64 → 128 GOPS across S7 Gen 1 → Elite Gen 2, explicitly to run multiple concurrent intelligent experiences (ANC adaptation, voice, context sensing) rather than one hero feature.
The sober engineering reading: these GOPS figures are tiny by phone standards (tens vs. tens of TOPS), so model architecture (small RNNs/CNNs, sparse activation, streaming inference) matters more than peak MACs; the NPU's real gift is energy per inference, enabling always-on keyword spotting and scene classification at sub-milliwatt averages.

7. Spatial Audio, Sensing, and the Health Turn

7.1 Head-Tracked Spatial Audio

The rendering split is a genuine architectural decision:
  • Phone-side rendering (Apple's model): the phone convolves HRTFs and streams ordinary stereo; buds return IMU pose over the control link. Keeps buds simple; costs motion-to-sound latency.
  • Bud-side rendering: Ambisonic/object audio streamed; buds convolve locally using on-die IMU fusion. Qualcomm positions Elite Gen 2's XPAN bandwidth + low-latency sensor compute as a "step change in head-tracking latency" for exactly this class of experience.
Either way the SoC contributes an always-on sensor hub fusing 6-axis IMU data (complementary/Kalman filters) at tens of hertz with sub-50 ms motion-to-sound targets, while the audio domain sleeps between ISO events.

7.2 In-Ear Detection and Interaction

Capacitive and IR proximity sensing, touch/force interfaces, and increasingly ML-based wear detection run in the always-on domain; their design goal is single-digit microamp average current while remaining responsive enough to pause audio within ~100 ms of removal.

7.3 Biometric and Hearing-Health Integration

  • PPG heart rate: AirPods Pro 3 embed Apple's smallest custom PPG sensor, pulsing IR 256 times per second, fused with accelerometer/gyroscope/GPS and an on-device AI model on iPhone for workout tracking; Powerbeats Pro 2 shipped the same capability earlier in 2025. This adds an optical analog front end and LED drive to the earbud's mixed-signal burden — a new SoC requirement class.
  • Hearing augmentation: The regulatory opening (US OTC hearing aids) plus LE Audio's Hearing Access Profile turned earbuds into clinical-adjacent devices. Apple now ships an end-to-end Hearing Test / Hearing Aid / Hearing Protection stack (with EU/UK certification for the latter). On the broadcast side, the SIG announced in October 2026 that all five major hearing-aid manufacturers now embrace Auracast — the moment assistive listening and consumer TWS became one interoperable ecosystem.
  • Design consequence: hearing-grade latency (<10 ms own-voice path), personalized amplification curves, and safe-listening dose monitoring are becoming standard DSP/NPU feature sets, not niche SKUs — Qualcomm explicitly markets dedicated cores for hearing-loss compensation and personal sound amplification.

8. Power Architecture and Energy Engineering

8.1 The Governing Arithmetic

Runtime T = (C_bat × V_nom × η) / P_avg Example: 60 mAh × 3.7 V ≈ 222 mWh; T = 8 h ⇒ P_avg ≈ 28 mW ≈ 7.5 mA @ 3.7 V
AirPods Pro 3's claimed 8 h with ANC on (a 33% generation gain) and 10 h in Transparency sit exactly at this envelope — evidence that flagship runtime gains now come from silicon/process and duty-cycle engineering, not bigger cells.

8.2 Techniques Stack

  1. Power domains: always-on island (RTC, sensor hub, KWS) separated from application, audio, and radio domains with independent retention and DVFS.
  2. Radio duty cycling: LE ISO event anchoring lets the radio sleep for most of each interval; Classic sniff/eSCO equivalents are cruder — a core LE Audio efficiency argument.
  3. Hardware offload: link controller, ANC biquad engines, and fixed-function codec paths run without waking the CPU.
  4. Process: the move to 6 nm FinFET (BES2800) cuts both dynamic and leakage power — decisive for ambient, always-worn use cases, which is why Qualcomm's headline Elite Gen 2 metric is a 0.6 mA Wi-Fi-connected idle simulation target.

8.3 Illustrative Power Budget (Engineering Estimates, Not Specifications)

Block
Music + hybrid ANC, modern silicon
Radio + link (LE Audio unicast, moderate retx)
~6–12 mW
DSP audio graph + ANC
~5–10 mW
DACs + earpiece drivers (level-dependent)
~3–6 mW
3–4 MEMS mics + ADC front ends
~2–4 mW
CPU, sensors, PMU losses, memory
~3–6 mW
Total
~20–38 mW → ~6–11 h on 222 mWh

8.4 The Case: The Forgotten Third Node

The case carries its own MCU + charger + boost converter (5 V pogo output), battery protection, and increasingly a BLE/UWB identity beacon for network-based finding. Bud–case communication over the charging contacts (or BLE) coordinates pairing, DFU staging, and state-of-charge display. Case-side innovation is now product-differentiating in its own right (e.g., premium cases with radios for precision finding).

9. Software, Security, and Production Engineering

9.1 Software Stack

  • RTOS-based firmware (vendor SDKs: Qualcomm ADK-lineage, Airoha SDK with tuning tools and reference iOS/Android/PC apps, Bestechnic SDK), with the audio graph expressed in framework form (Kymera being the most explicitly documented).
  • Dual-bud atomic OTA: A/B images with rollback, cross-bud version lockstep, and case-staged transfers — a genuinely distributed-systems problem at 100 kB/s-class control bandwidths. Airoha lists FOTA as a first-class SDK feature.
  • Ecosystem hooks: Google Fast Pair, Microsoft Swift Pair, LE Audio CSIP set provisioning; proprietary layers (Apple pairing/Find My, Huawei/HarmonyOS integration).

9.2 Security

The threat catalog is mature: KNOB (entropy negotiation), BIAS (impersonation), BLURtooth (cross-transport key derivation), and keystroke-injection classes (CVE-2023-45866). Mitigations live in the controller/host boundary: Secure Connections-only policies, key-size floors, cross-transport key isolation, and patched link managers. TWS adds its own surface — mirroring requires link-key sharing between buds, and Auracast adds Broadcast_Code management. Silicon answers with secure boot chains, OTP-rooted keys, TrustZone-M partitioning, and debug lockdown; counterfeit detection/provisioning is a live commercial concern at the value tier.

9.3 Manufacturing and Calibration

Per-unit factory flows are where ANC performance is actually made: microphone sensitivity/phase matching, driver response capture, ANC filter trimming, crystal trim, and RF calibration — followed by acoustic end-of-line testing. Regulatory overlays include SIG qualification, regional RF certification, loudness standards (EN 50332 class), and, for hearing features, medical-device regimes (FDA OTC; EU MDR) — Apple's regional gating of hearing features illustrates the compliance burden.

10. Competitive Landscape and Strategic Dynamics

Vendor
Representative silicon (verified)
Position
Differentiation
Qualcomm
S3/S5 Gen 3
; S7/S7 Pro Gen 1; Elite Gen 2
Premium merchant
Full-stack Snapdragon Sound; XPAN Wi-Fi; eNPU; ANC IP; Bose/HP/Cleer design wins announced
Airoha (MediaTek)
AB1562; AB157x; AB1585; AB1595
Volume-to-flagship merchant
Aggressive integration (flash, PMU, AiroHAP); voice/ANC tooling; MediaTek HDR link
Bestechnic
BES2800/BES2800BP
Android-OEM wearables
6 nm leadership; BT+Wi-Fi 6 combo (Ceva IP); glasses/wearable expansion
Realtek / Bluetrum / JieLi / Actions
RTL876x/877x; AB56xx; AC7xxx; ATS
Value tier
Cost, RISC-V, turnkey firmware; ANC commoditization
Apple
H2 (AirPods Pro 3)
Captive
Vertical co-design: silicon + OS + services; hearing health; PPG
Huawei
Kirin A2 (FreeBuds Pro 4); Kirin A3 (FreeBuds Pro 5)
Captive
NearLink + L2HC 2.3 Mb/s lossless; HarmonyOS integration
Google
Tensor A1 (Pixel Buds Pro 2 / 2a)
Captive
Dedicated audio pipeline; Gemini/assistant integration
Sony
V2 + QN2e (XM5) / QN3e (XM6, Feb 2026)
Captive/co-designed
Two-chip ANC offload; LDAC; audio-brand tuning
Samsung
Merchant SoC + SSC UHQ codec
Hybrid
Galaxy-ecosystem codec; LE Audio/Auracast early adoption
Dynamics worth tracking:
  1. LE Audio commoditizes the baseline; differentiation migrates upward to AI voice, adaptive ANC, and multi-radio connectivity
  2. Standardization is absorbing proprietary advantages — HDT/lossless profiles (Nov 2026) do to aptX Lossless/L2HC what LE Audio did to mirroring.
  3. Hearing health is the new regulatory moat — clinical validation and regional certifications favor vertically integrated players, while HAP/Auracast pull the whole industry along.
  4. The phone-tether is loosening: direct-to-cloud earbuds (XPAN + apps/agents) reposition the bud as an ambient computing endpoint.

11. Outlook: 2026–2028

  1. HDT productization. First standardized lossless-music earbuds over Bluetooth 7.0 HDT; expect ~2027 devices from Elite Gen 2-class silicon, with graceful LE 2M fallback under congestion.
  2. Channel Sounding everywhere. Case-finding and proximity UX without UWB companions, hardened by 6.2/6.3 anti-attack and accuracy features.
  3. Wi-Fi/Bluetooth coexistence as a standards project. The SIG and Wi-Fi Alliance have initiated joint coexistence work — a prerequisite for Wi-Fi-equipped hearables at scale.
  4. Agentic hearables. On-bud assistants with cloud fallback, real-time translation (already shipped in constrained form by Apple), and camera-equipped "context buds" — Qualcomm explicitly targets multimodal, camera-enabled form factors with privacy-conscious sensing.
  5. Energy for ambient wear. Sub-milliamp connected-idle targets plus 6 nm-class leakage enable all-day worn, occasionally-used devices — the open-ear/glasses convergence.
  6. Compute realism. 128 GOPS-class NPUs will host small streaming models, not LLMs; the LLM stays on phone/cloud, and the earbud's AI value remains in the acoustic front end.

12. Engineering Takeaways

  1. A TWS SoC is best understood as a microsecond-latency control system (ANC), a millisecond-latency streaming system (audio), and a second-latency policy system (AI/UX) sharing one milliwatt budget.
  2. Topology is destiny: the relay→mirroring→CIS/BIS progression is the single most consequential protocol history in the category; LE Audio made binaural synchronization a standards problem with a standards solution.
  3. Causality and Bode, not marketing dB, bound ANC; published 3 µs DSP paths and 112-biquad engines show where the real engineering lives.
  4. The codec ceiling is an RF ceiling; HDT's 7.5 Mb/s is the first standards-based escape, while XPAN/NearLink previewed the multi-radio future.
  5. Heterogeneity won: M-class control + HiFi DSP + dedicated low-latency core + NPU is now the canonical compute template across Qualcomm, Airoha, and Bestechnic.
  6. The category's center of gravity is shifting from "wireless headphones" to worn, sensing, hearing-health, AI-interface devices — and the silicon roadmap (integrated Wi-Fi, sub-mA idle, PPG front ends, secure ranging) reads as evidence, not speculation.