Dolphin Design – Delivering High-Performance Audio Processing with TSMC’s 22ULL Process

TSMC held their very popular Open Innovation Platform event (OIP) on August 25. The event was virtual of course and was packed with great presentations from TSMC’s vast ecosystem. One very interesting and relevant presentation was from Dolphin Design, discussing the delivery of high-performance audio processing using TSMC’s 22ULL process through their computing platforms and subsystems.

The OIP event followed TSMC’s Technology Symposium, which was held the day before. I’ve heard from more than one person that these virtual events were well produced, easy to follow and had the added advantage of not needing to get up at the crack of dawn to get a parking spot and a good seat. Virtual events are clearly the new normal.

Dolphin’s presentation began by discussing the business trends for AI applications in audio markets. This was followed by a discussion of ultra-low power (uLP) audio processing, an application use case and an overview of Dolphin’s platforms for audio processing. I’ll provide some highlights of each section of their presentation here.

Business Trends in AI Audio Markets

This section began by pointing out that voice is the easiest form of a user interface. This includes the following properties:

• Intuitive
• Quick and accurate
• No contact
• Straightforward
• Easy integration

Voice-enabled devices need to address several technical challenges, including:

• Voice detection
• Keyword spotting
  • Voice pickup & noise reduction
• Speaker separation
• Active noise control
• Speech recognition
• Low power

So, voice-enabled devices represent the next revolution for user experience. The opportunity is to provide power optimized, local AI processing for things like speech recognition, wake-word detection and voice detection. Local processing will deliver better latency, lower cost and improved privacy since voice data is not sent to the cloud.

uLP Voice Detection and Keyword Spotting

Dolphin Design provided some very good detail on the benefits of their IP and associated platforms for voice detection. You can also see Tom Simons’s post on Dolphin Design and voice detection here. The figure below illustrates the high-performance and ultra-low power audio processing they can deliver for voice detection.

The Dolphin approach for voice detection provides the following benefits:

• Stand-alone IP embedding a smart algorithm to detect voice activity
• Automatic tuning of detection algorithms to the level of background noises
• Short detection latency to avoid the need of buffering the audio stream
• Ambient noise sensing for optimal adaptation of the key word spotting (KWS) algorithm to environmental conditions

A typical record lifetime of systems with a 25 mAh battery is ~5 hours without Dolphin technology and ~38 hours with Dolphin technology.

For keyword spotting, Dolphin Design can also deliver high-performance and ultra-low power audio processing using their MCU subsystem as shown in the figure below.

Using Dolphin’s CHAMELEON MCU subsystem yields the following benefits:

• Up to 80x power reduction
• Bringing KWS in µW range
• No need for accelerator
• Enables faster inference
  • for multiple speakers
  • for beamforming
  • still in mW range

Application use case: True Wireless Stereo (TWS) Earbuds

An example application for TWS earbuds was presented. Several Dolphin Design platforms and subsystems were used in this application. The benefits of each of these capabilities can be summarized as follows:

• CHAMELEON MCU Subsystem
  • Compatible with main MCU
  • High bandwidth through low latency interconnect
  • Tiny ML accelerator with 32 MAC/cycle
  • <20 µA/MHz & 2µA deep sleep in TSMC 22uLL

• BAT Audio Platform
  • Up to 768 kHz sample rate
  • Less than 7us analog to analog latency
  • Up to 8 analog and digital mic inputs
  • I2S/AHB data interface & I2C/APB control interface

• SPIDER Power Management
  • Customizable & tailored power network
  • Standardized & predictable power management
  • 250 nA quiescent DCDC
  • 150 nA quiescent LDO

• PANTHER DSP
  • Up to 64 MAC/cycle
  • Up to 16 cores scalability
  • Standard AXI interface
  • Enhanced SIMD DSP, NN instructions

Dolphin Design Platforms for Audio Processing

The following diagram summarizes Dolphin Design platforms and their capabilities in the field of audio and processing applications.

Dolphin summarized how they are delivering high-performance audio processing with TSMC’s 22ULL processes follows:

• Audio/Voice markets will be dominant AI market in coming years
  • Smart Sensors approach will be the driving force
• Dolphin Design has a long experience in Audio Codecs
• New platforms will enable Voice User Interface
  • uLP speech recognition for enabling the voice-control world
  • Open platform as a design Backbone reusable for multiple projects, multiple processes, multiple processor vendors
  • Reduce key expertise bottlenecks
  • Faster TTM thanks to ready-to-use audio platform

You can learn more about the platforms and systems available from Dolphin Design here