Mobile Audio Subsystem: DAC, Amplifier and Spatial Audio Architecture

If you think smartphone audio is just a DAC and a speaker, you're roughly fifteen years behind. Today's flagship audio subsystem is a heterogeneous signal chain spanning three clock domains, two operating systems, and a dedicated neural processing path — all optimized for 3mW idle power.

On a recent Pixel 9 Pro teardown, the Google Tensor G4 audio island occupied less board area than the haptic driver, yet it delivered 123dB dynamic range, sub-5µs inter-channel skew, and real-time Dolby Atmos virtualization for two speakers spaced 14cm apart. That is not a codec. That is a compute subsystem. Understanding its architecture explains why some phones sound dramatically larger than their physics should allow.

1. The Signal Chain: From PCM to Pressure

A mobile audio path is best viewed as five stages with very different constraints: source/mixing, DSP processing, conversion, amplification, and transduction. Android's AudioFlinger (or Core Audio in iOS) mixes 32+ streams at 48kHz/24-bit, applies volume curves, then hands a single interleaved PCM buffer to the Audio HAL. From there, the path leaves the AP and enters the audio subsystem, usually a dedicated island with its own SRAM, PLL, and power rail.

This island exists for one reason: determinism. The application processor (AP) cannot guarantee microsecond-level jitter when it is also rendering a game at 120Hz. The audio DSP — Qualcomm's ADSP, Exynos' ABOX, or Apple's Always-On Processor — runs a real-time OS (QURT, FreeRTOS) that guarantees 1-2ms scheduling latency. All sample-rate conversion, dynamic range compression, and spatial rendering happens here, not on the AP.

2. DAC Architecture: Integration vs. Performance

Five years ago, audiophiles mourned the removable 3.5mm jack because it often removed the discrete DAC. Today, the story is more nuanced. The integrated CODEC inside Snapdragon 8 Gen 3 (WCD9395), Tensor G4, and Dimensity 9300 is objectively excellent: 32-bit delta-sigma modulators, 123-125dB(A) dynamic range, -105dB THD+N, built on 6nm analog-friendly process with dedicated MIM caps.

So why do LG V60-style discrete ESS Sabre designs still measure better? Two reasons: power isolation and analog voltage headroom. An integrated DAC shares substrate with screaming digital logic. Even with deep N-well isolation and separate LDOs, you get ~2-3dB higher noise floor than a discrete Cirrus Logic CS43131 or ESS ES9281AC fed by a clean 1.8V rail. Discrete parts also run at 2Vrms (vs 1Vrms integrated) which better drives high-impedance loads without an external amp.

That said, for 95% of users on 32Ω IEMs or Bluetooth, the integrated DAC is transparent. Modern delta-sigma designs oversample at 384kHz with 5th-order noise shaping, pushing quantization noise to ultrasonic bands where a simple RC filter removes it. The audible bottleneck is rarely the DAC — it is the amplifier and transducer. This is why smartphone BOM analysis now allocates 3× the audio budget to smart amplifiers versus DACs compared to 2019.

The 3.5mm Question

When a phone retains a jack, it typically adds a dedicated headphone amp with programmable gain (0-12dB) and impedance sensing. The CODEC detects 16Ω to 600Ω loads and adjusts charge-pump voltage to optimize THD. On USB-C audio, the DAC moves to the dongle — a $9 dongle with a decent CS43131 can outperform a $999 phone's internal path simply because the dongle's ground is cleaner. For wired audiophile use, the dongle is now the reference path.

3. The Amplifier Stage: Where Watts Meet Millimeters

Smartphone speakers are a physics paradox: 11×15mm micro-speakers with 0.6mm excursion trying to reproduce 80Hz. Without intelligence, they would either sound thin or self-destruct. The modern solution is the smart amplifier — a Class-D amplifier with integrated IV sense, DSP, and protection model.

Parts like the Cirrus Logic CS35L56, Texas Instruments TAS2563, and Qualcomm WSA8835 output 5.1-5.5W into 8Ω from a 5V supply boosted to 11V via an internal DC-DC. But peak power is not the story. The amplifier continuously measures voltage and current (IV) at 48kHz, feeds that into a thermal model and an excursion model, and predicts voice-coil temperature and diaphragm position 1-2ms ahead.

This allows two tricks:

• Thermal overdrive: Play 2× RMS power for 200ms transients because the model knows the coil can absorb it without exceeding 110°C. Results in dramatically punchier drums.
• Excursion protection: Apply dynamic low-shelf filtering that reduces bass only when predicted excursion exceeds Xmax, rather than a static high-pass. You get extended bass at low volume, safety at high volume.

Class-G and Class-H variants dynamically switch supply rails to save power. At typical listening levels (~80dB SPL at 0.5m), a Class-G amp is 25-30% more efficient than pure Class-D because it idles on a 2.5V rail and only jumps to 5V or 8V on peaks. That efficiency pays for itself in battery — audio playback is now dominated by speaker efficiency, not amplifier overhead. See our analysis of wireless charging coil design for how shared magnetics and tight Z-height constraints force amp placement directly under the speaker enclosure.

4. Spatial Audio Architecture: Two Speakers, Infinite Positions

Spatial audio on mobile is not about adding speakers. It is about psychoacoustics — convincing the auditory cortex that sound comes from outside the device. Dolby Atmos, Sony 360 Reality Audio, and Apple's Spatial Audio all use object-based rendering, where each sound source has metadata: azimuth, elevation, distance, size, and occlusion.

The rendering pipeline on Snapdragon's Hexagon or Apple's H1-derived spatial core works as follows:

1. Decode: Atmos DD+JOC or MPEG-H bitstream → up to 16 objects + 7.1.4 bed.
2. Prerender to binaural: Apply Head-Related Transfer Function (HRTF) filters (typically 128-tap FIR per ear at 48kHz) based on generic or personalized HRTF.
3. Crosstalk cancellation: For speaker playback, invert the acoustic path between left speaker→right ear to widen stereo image. Requires precise inter-speaker delay calibration (±10µs).
4. Room simulation: Early reflections (6-12 taps) + late reverberation (feedback delay network) to add out-of-head localization.
5. Head tracking: Fuse IMU (gyro + accel) at 200Hz to rotate soundfield opposite to head movement, latency <20ms to avoid motion sickness.

On phones without head tracking (most Android), the renderer locks the soundstage to the device — still effective for movies because visual anchoring dominates. For headphones, the phone either renders and streams binaural PCM over Bluetooth LE Audio (LC3+) or sends objects over Bluetooth and lets the earbuds render (lower latency, personalized HRTF on earbud DSP).

The biggest challenge is not compute — a 16-object Atmos renderer costs ~120 MIPS, trivial for a modern DSP. It is tuning. Each phone's speaker response, enclosure resonance (typically 800Hz peak), and acoustic leakage through hands must be compensated. OEMs ship per-device EQ and delay tables generated via automated end-of-line acoustic calibration. This is why two phones with identical silicon can sound completely different.

5. Power, Latency, and the Thermal Envelope

Audio power looks small until you multiply by time. Streaming Spotify via speakers at moderate volume draws 600-900mW system power, of which 250-400mW is the amplifier, 40mW the DSP, and 8mW the DAC. For always-on voice wake ("Hey Google"), the DSP + mic bias must stay under 5mW or standby battery collapses. That is why voice wake runs on a tiny Cortex-M55 or HiFi 3z core at 50MHz with its own 32kHz clock — completely decoupled from the main AP.

Latency targets diverge by use case. Music playback tolerates 50ms buffering. Voice calls need <20ms round-trip to feel natural. Gaming and spatial video need <10ms A/V sync. Android's MMAP/AAudio low-latency path bypasses the mixer and talks directly to the DSP via shared memory, achieving 4-6ms output latency on modern HALs — down from 20ms in 2019. Bluetooth remains the laggard at 80-120ms, which is why LE Audio with LC3 and iso-periodic intervals targeting 20-30ms is so critical.

6. How to Measure and Compare

Specs lie if you don't know the load and level. Always demand conditions: THD+N at 1kHz, 32Ω, 1Vrms, A-weighted, and crosstalk with both channels driven. The table below compares representative designs from teardown measurements.

* Speaker THD dominated by micro-speaker nonlinearity, not amplifier. Measured at 1W in 11×15mm enclosure at 1kHz.

For software developers, the key takeaway from mobile ML framework comparisons applies here too: the hardware provides headroom, but model tuning (EQ, DRC, HRTF personalization) determines perceived quality. A 6-band parametric EQ tuned by ear beats a 130dB DAC with flat tuning.

7. What's Next: Lossless Wireless and Personal HRTF

The next frontier removes the final compromise: wireless. Bluetooth LE Audio with LC3plus at 345 kbps is approaching perceptually lossless for most listeners, and Qualcomm's XPAN and Apple's H2 proprietary modes already stream 24-bit/48kHz lossless over Wi-Fi to earbuds when in range. On the capture side, 24-bit/192kHz recording with active noise cancellation mic arrays is standard.

More exciting is personalized spatial audio. iOS and recent Snapdragon phones scan ear geometry via FaceID / selfie camera to select or morph an HRTF from a 200+ person database, improving elevation accuracy from ~30° to ~12° median error. Research prototypes using in-ear mics for acoustic HRTF measurement promise <8° error — but require user calibration. Expect this to become a setup step like display calibration.

Until then, the fundamentals remain: isolate the DAC supply, give the amp a model of its speaker, and render spatial cues with microsecond timing. Phones that get this right — regardless of brand — sound disproportionately large. Those that treat audio as a checklist item sound small, noisy, and flat, no matter how many drivers they advertise.