Behind the smart speaker: The "sound chip" family of Rockchip
From RK3308 to RK2118G: Why Don't Smart Speakers Choose General-Purpose SoCs?
Your home's Xiao Ai, Tmall Genie, or Xiaodu speaker—when you take it apart, it's highly likely not powered by a "versatile" chip like the RK3568, but rather a dedicated audio SoC.
Why? Because using a general-purpose SoC for speakers presents three problems:
1. Audio interface is insufficient - I2S, PDM, TDM, SPDIF, HDMI ARC, universal. For general SoC, usually only 1-2 channels of I2S are provided.
2.Power consumption is too high - even the four-core A55 processor consumes several hundred milliwatts in standby mode, and the speaker is online 24/7.
3.No VAD hardware acceleration - Voice Activity Detection (which detects whether you are speaking) is all handled by software, resulting in high latency.
The answer from Rockchip is: The RK3308 series and the RK21xx series.
一、RK3308 Family: Quad-core A35 Audio Expert
Chip | CPU | Memory | Feature |
RK3308 | 4xA35 | 256KB SRAM | Standard version, external DDR required |
RK3308B | 4xA35 | 256KB SRAM | CAN interface removed |
RK3308GK | 4xA35 | 64MB DDR2 Integrated | No external memory required |
RK3308H1 | 4xA35 | 128MB DDR3L Integrated | Larger memory, plug-and-play |
RK3308H1-S | 4xA35 | 128MB DDR3L | The small-sized version of H1 |
RK3308H2 | 4xA35 | -- | variant |
RK3308H3 | 4xA35 | -- | Latest variant (2025) |
RK3308J | 4xA35 | -- | Industrial grade (wide temperature range) |
RK3308M | 4xA35 | -- | Medical grade |
RK3308K | 4xA35 | -- | Special packaging variant |
Core Specifications
How Impressive Is the Audio Interface?
The Audio Interface Configuration of RK3308
•2x I2S_2CH(2 channels per lane)
•4x I2S_8CH(It can be aggregated as I2S_16CH, while I2S_8CH_3 only supports the 4CH RX mode.)
•PDM 8ch(Connect directly to the PDM microphone)
•TDM 8ch(Time Division Multiplexing)
•SPDIF TX/RX(Optical fiber audio)
•HDMI ARC(Audio transmitted from the television)
What Does This Mean?
A Single Chip Can Simultaneously Connect:
•6 to 8 microphones (for far-field sound pickup)
•Two speakers (stereo)
•1 optical fiber audio input/output
•1 HDMI ARC (for taking audio from the TV)
No need for any external audio expansion chips.
VAD Hardware Acceleration
VAD(Voice Activity Detection)This is the core function of the smart speaker: to detect whether the user is speaking and determine when to activate voice recognition.
RK3308 Uses Hardware for VAD Instead of Relying on the CPU - The Benefits:
•Extremely low standby power consumption: The CPU can enter sleep mode during VAD detection.
•Lower latency: Hardware detection is several milliseconds faster than software detection.
•Power-saving: A speaker that can remain in standby mode for 7×24 hours. This is crucial.
RK3308H1: Embedded 128MB DDR3L
The biggest difference between RK3308H1 and other versions: the memory is built-in.
Comparison Item | RK3308 Standard Version | RK3308GK | RK3308H1 |
Built-in memory | 256KB SRAM | 64MB DDR2 | 128MB DDR3L |
External memory | need(DDR2/DDR3) | NO | NO |
PCB Area | Large (with external DDR chip) | little | little |
BOM Cost | High (with one more DDR) | low | low |
I2S | 2x I2S_2CH + 4x I2S_8CH | Same as above | Same as above |
显示 | 1080p RGB/MCU | 1080p | 1080p(+gamma/dithering) |
Note: The I2S configuration is identical across all RK3308 variants (including the standard version): 2 × I2S_2CH + 4 × I2S_8CH, which can be aggregated into 16CH. The real advantage of H1 lies in larger onboard memory (128MB DDR3L vs 64MB DDR2).
Applicable Scenarios
Scene | Recommended model | Reason |
Smart Speaker (External DDR Solution) | RK3308 / RK3308B | Standard Version, Flexible Configuration |
Smart Speaker (Lite Solution) | RK3308H1 | Onboard 128MB, Single-chip Solution |
Low-Cost Speaker | RK3308GK | Onboard 64MB, Lowest Cost |
Industrial Environment Speaker | RK3308J | Wide Temperature Certification |
Display Smart Speaker | RK3308H1 | VOP Supported 1080p + gamma/dithering |
RK21xx Series: MCU-class Ultra-Low Power Audio
The RK3308 is a Linux-class SoC (Quad-core A35, capable of running Android/Linux). However, certain use cases do not require such powerful CPU performance — for example, compact devices that only need voice wake-up and audio playback.
This is where the RK21xx series comes into play.
Family Members
Chip | CPU | DSP | NPU | Onboard Memory | Positioning |
RK2106 | 2xCortex-M3 | -- | -- | ~1MB RAM | Wireless Audio/ MP3/IoT |
RK2108 | Cortex-M4F (+FPU) | HiFi3 DSP | -- | ~1MBShared Memory | Smart Home / IoT / Voice Applications |
RK2108C | M4F | HiFi3 DSP | -- | -- | RK2108 Variants |
RK2108D | M4F | HiFi3 DSP | -- | -- | RK2108 Variants |
RK2116G | 2xStar-SE (ARMv8-M) | 3xHiFi4 DSP | Audio NPU | 1024KB + Embedded DDR2 | Smart Voice / AI Audio |
RK2116M2 | Star-SE + HiFi4 | 3xHiFi4 | Audio NPU | 1024KB + Embedded DDR2 | M.2 Package |
RK2118G | 2xStar-SE (ARMv8-M) | 3xHiFi4 DSP | Audio NPU | 1024KB +Embedded DDR2 | Smart Voice / AI Audio |
RK2118M | Star-SE + HiFi4 | 3xHiFi4 | Audio NPU | 1024KB +Embedded DDR2 | M Package |
RK2118M2 | Star-SE + HiFi4 | 3xHiFi4 | Audio NPU | 1024KB +Embedded DDR2 | M.2 Package |
RK2206 | Cortex-M4 (+DSP) | -- | -- | Internal SRAM | IoT / Wireless Audio |
RK2106: The Lowest-Cost Audio SoC
Item | Specification |
CPU | 2xCortex-M3 |
Memory | 64KB PMU SRAM + 320KB SYSRAM0 + 256KB SYSRAM1 + 128KB HIGHRAM0 + 256KB HIGHRAM1 ≈ 1MB |
Audio | 2×I2S (Up to 192KHz), Codec ADC |
Display | i8080 MCU interface, up to 400×400 |
Features | Hardware MP3 Decoder Accelerator (IMDCT36 + sub‑band synthesis) |
Power Consumption | 3 power modes, ultralow standby power |
Positioning | MP3 Player, Wireless Audio, IoT |
The key selling point of the RK2106 is that no external memory is required — all RAM is integrated within the chip. Combined with the built-in hardware MP3 decoder, an MP3 player can be implemented with just this single chip, Flash and battery.
RK2108: M4F + HiFi3 DSP
Item | Specification |
CPU | Cortex-M4F(带 FPU),16KB I-Cache + 16KB D-Cache |
DSP | HiFi3 DSP (4×24‑bit MAC / Dual 32‑bit MAC, 3 VLIW slots, 2‑Way SIMD VFPU) |
DSP Memory | 64KB/512KB I/D TCM |
Audio | I2S (2ch+6ch)、PDM 6ch、Codec ADC、Audio PWM |
Video | VIP Input (BT601/BT656), VOP up to 800×480 |
Features | VAD, Audio Bypass Path, 4 power modes, Automatic DSP power reduction |
Positioning | Smart Home, IoT, Voice Interaction |
The HiFi3 DSP is a licensed core from Cadence and an industry benchmark in audio processing. It can efficiently run various audio algorithms.
Acoustic Echo Cancellation (AEC)
•Noise Suppression (NS)
•Beamforming
•Speech Enhancement
RK2118G: Top-Tier Audio SoC
Item | Specification |
CPU | 2×Star‑SE (ARMv8‑M Architecture), FPU, TrustZone |
DSP | 3xHiFi4 DSP(4 VLIW, 64-bit SIMD, 4x32x32 MAC) |
DSP Memory | DSP0: 256KB ITCM + 768KB DTCM,DSP1/2: 64KB ITCM + 256KB DTCM |
NPU | Audio NPU (32×FP16 MAC/cycle), supports TensorFlow / Caffe / TFLite / PyTorch / ONNX |
Audio | 5×SAI, PDM, Dual ASRC, FIR/IIR Hardware Accelerator (4096tap FIR, 64 biquad IIR) |
Display | VOP, up to 480×480 (i8080 MCU Interface) |
Interfaces | USB 2.0 OTG、RMII、CAN、FSPI、eMMC、SD、2xSPI、4xUART、I2C、FLEXBUS |
Memory | Embedded DDR2 + 1024KB System RAM |
Positioning | Smart Voice Interaction, Audio AI Processing, Smart Speakers |
RK2118G Standout Specifications:
1、CPU: 2xStar-SE (ARMv8-M)
2、DSP: 3×HiFi4 (Professional Audio DSP)
3、Audio‑Dedicated NPU (FP16, supports ML model inference)
4、A complete AI audio processing platform
Note: The RK2118G does not support PCIe or USB 3.0; only USB 2.0 OTG is available. Display output is via the i8080 MCU interface with a maximum resolution of 480×480 (not 1280×1280).
This chip can run AI audio models locally. Keyword spotting, voice recognition front-end, noise classification and acoustic scene recognition are all hardware-accelerated by the NPU, with minimal CPU involvement.
RK2116 vs RK2118
Comparison Item | RK2116G | RK2118G |
CPU | 2xStar-SE | 2xStar-SE |
DSP | 3xHiFi4 | 3xHiFi4 |
NPU | Audio NPU | Audio NPU |
Memory | 1024KB + DDR2 | 1024KB + DDR2 |
Differences | M.2 Package, Suitable for Module Design | Standard Package, Suitable for Mainboards |
RK2116 and RK2118 share nearly identical silicon die; the difference lies in packaging. The RK2116 adopts the M.2 module form factor (pluggable like a Wi-Fi card), while the RK2118 comes in a standard package.
RK35xx Family Audio Specialist
Apart from the RK3308 and RK21xx families, the RK35xx series also includes SoCs targeted at audio applications.
RK3502G2: Robot Vacuum Cleaner
Item | Specification |
CPU | 3xCortex-A7,128KB L2 |
GPU | None (Only 2D Engine, up to 1280×1280) |
Memory | On-chip 128MB DDR3L |
Audio | 5xSAI、PDM 8ch、ASRC(2x)、Audio DSM、SARADC |
Coprocessor | Cortex-M0 |
Positioning | Robot Vacuum Cleaners |
Why do robot vacuum cleaners adopt audio SoCs? They require voice prompts ("Start cleaning", "Low battery") and voice commands ("Return to charge"), alongside running motion control algorithms.
Triple Cortex‑A7 cores deliver sufficient performance. The 5×SAI interfaces connect to microphone arrays, and the integrated 128MB DDR3L helps save PCB space.
RK3506B/G1/J: Smart Speakers / Display Devices
Item | RK3506B | RK3506G1/G2 | RK3506J |
CPU | 3xA7 | 3xA7 | 3xA7 |
Memory | External DDR2 / DDR3 | Embedded 64MB DDR2 | External DDR2 / DDR3 |
Audio | 2xSAI, PDM, SPDIF | 5xSAI, PDM, ASRC, DSM, ADC | SAI, PDM |
Display | 2xMIPI DSI | MIPI DSI | -- |
Special Interfaces | CAN, RMII | CAN, RMII, FLEXBUS | CAN, RMII |
Positioning | Smart Speakers / Voice Applications | Smart Displays / Voice Applications | Industrial / Automotive |
The RK3506 series is an ideal choice for smart display speakers: 3×Cortex‑A7 cores run lightweight Linux, supporting small-screen display and comprehensive audio interfaces.
二、RK730 Audio Codec
Item | Specification |
Type | Pure Audio Codec Chip |
Functions | ADC + DAC, Headphone Output, Speaker Output, Differential Mode |
Positioning | Portable Audio Devices |
The RK730 is not an SoC, but a dedicated pure audio codec chip. If your main controller (such as the RK3568) lacks sufficient audio interfaces or requires higher-quality audio output, the RK730 can be used as an external chip
二、Horizontal Comparison: Voice SoC Selection Matrix
Application Scenario | Recommended Chip | Reason |
Low-cost Smart Speakers | RK3308GK | Embedded 64MB memory, single-chip complete solution |
Premium Smart Speaker Solution | RK3308H1 | Embedded 128MB memory, most comprehensive interfaces |
Smart Display Speakers | RK3506G1 | 3×CortexA7, 64MB embedded memory, MIPI DSI supported |
Voice Remote Controls | RK2106 | Dual-core M3, ultra-low power consumption |
Voice Interaction IoT Devices | RK2108 | M4F + HiFi3 Audio DSP, hardware VAD support |
Local AI Audio Recognition | RK2118G | Dedicated Audio NPU, supports TensorFlow / PyTorch model deployment |
Robot Vacuum Cleaners | RK3502G2 | 128MB embedded memory, 5×SAI, CortexM0 coprocessor |
Industrial Voice Terminals | RK3308J | Wide temperature range and abundant audio interfaces |
Summary
Rockchip’s strategy for audio SoCs is clear:
1.RK3308 Series: The preferred Linux-based solution for smart speakers. Equipped with abundant I2S interfaces and hardware VAD. The H1 variant integrates 128MB DDR for a highly simplified system design.
2.RK21xx Series: Ranging from M3 cores to Star-SE + HiFi4 + NPU architecture, covering full audio scenarios from basic MP3 playback to edge AI audio processing.
3.RK3502 / RK3506: General-purpose SoCs enhanced with rich audio interfaces, targeted at dedicated applications including robot vacuums and display smart speakers.
ONE MORE THING
Thinkcore Team is an embedded board solution provider, developer, and motherboard supplier, specializing in the design, manufacturing, sales, and technical sharing of embedded motherboards. Our chipset portfolio includes Rockchip, Allwinner, and others. We welcome inquiries and collaboration! E-mail: developer@thinkcore.cn
This article is reprinted from here
https://mp.weixin.qq.com/s/9n7RArulsPCdJyDlS5jy0A