Several converging trends are driving the need for low-energy audio/voice capabilities in a small form factor, especially in hearables and wearables. Whether the target consumer devices need to be always listening, playing back music, providing context awareness, as well as providing on-device artificial intelligence (AI), this requires not just an energy-efficient DSP but also the ability to add an AI engine.
To address this, Cadence Design Systems has introduced a new device in its Tensilica HiFi DSP family, the HiFi 1, a smaller version of its HiFi 3 DSP that offers between 11 to 16% lower area, 60 to 73% greater cycle and energy efficiency for ML-based “OK Google” keyword spotting and person detect applications, and greater than 18% cycle efficiency and 14% energy efficiency for LC3 decoding.
Prakash Madhvapathy, product marketing director for Cadence, emphasized to embedded.com that the new HiFi 1 is not just a cut-down version of the HiFi 3, but really a ground-up design to address energy-efficiency and performance requirements of always-on expectations from consumers. That means that it is not just for hearables and wearables which are going to be the initial key markets for the HiFi 1, but also other use cases which require always-on sensor fusion to run at very low energy levels.
Hence the HiFi 1 DSP features ultra-low energy consumption to extend the duration of voice communication and music playback, allowing always-on listening to voice commands with minimal impact to battery life, according to Cadence. This enables small form factor, low-cost consumer and mobile devices, as well as automotive and industrial devices, to offer increased functionality with low energy consumption.
This is aimed at a number of trends in the consumer electronics space. Increasingly, hearables and wearables are adopting the low complexity communications codec (LC3) standardized by the Bluetooth Special Interest Group (SIG) in 2020. In addition, consumer preference for hands-free and touch-free control is driving demand for always-on, or always-listening, devices that respond to voice wake-up commands, and appliances are even adopting these capabilities.
To enable these devices and applications, Madhvapathy commented, “The name of the game for HiFi 1 is battery life for applications where energy is important.” He said that meant doing things like creating special instructions for arithmetic coding of LC3 encode and decode; this makes HiFi 1 the most energy-efficient DSP for Bluetooth LE audio. Other features of the new optimized HiFi 1 DSP include:
- Instructions for ITU-T3GPP-standardized 2019 BASOPs, to accelerate speech codecs and run them with high energy efficiency, increasing talk time.
- Neural network ISA and load/store accelerate keyword spotting and other machine learning workloads for reduced energy consumption, while memory access-optimized ISA (instruction set architecture) improves performance with small cache sizes.
- Efficient signal processing to reduce energy and cycles for audio and speech pre- and post-processing.
- Optional, low-latency vector floating point unit (VFPU) to deliver higher FP throughput with lower energy consumption.
- Vector Boolean register to improve energy efficiency for conditional code.
Mike Demler, senior analyst at The Linley Group, set the scene on why something like this new DSP was needed. He said, ““Intelligent context-aware, always-on processing is becoming key for next-generation hearables to offer users a better auditory experience with greater utility and convenience. A blend of machine learning and traditional DSP algorithms is pervading these applications, yet battery life and form factor cannot be compromised. Cadence has established a strong leadership position with its widely adopted and supported Tensilica HiFi DSPs for audio, voice and speech applications. The new HiFi 1 DSP goes a step further by enabling these emerging target applications in a compact design—deftly melding AI, codec and DSP performance at ultra-low energy levels for energy-conscious designs.”
Fraunhofer, which co-invented LC3, said that LC3/LC3plus help minimize energy consumption for battery-constrained Bluetooth devices, and that Cadence had worked with Fraunhofer to optimize the codec on its Tensilica HiFi DSP over the years. Manfred Lutzky, head of the audio for communications department at Fraunhofer IIS, said, “We’re pleased to see them building upon that experience with the new HiFi 1 DSP that is energy- and cycle-optimized for LC3 and LC3plus. The HiFi 1 DSP embodies the codec’s unprecedented ultra-low energy levels, which should bring relief to small battery hearable and wearable devices with added features and help accelerate the widespread adoption of LC3 and LC3plus.”
Cadence had also collaborated with Google on TensorFlow Lite Micro (TFLM). Pete Warden, technical lead of TFLM at Google, said, “The HiFi 1 DSP from Cadence significantly lowers the energy required to run the always-on class of AI applications, such as the TFLM networks, speech wake word and person detect, enabling battery-constrained devices to run for longer lengths of time. We are excited to continue the collaboration as Cadence pushes the boundaries of energy and performance further.”
Madhvapathy said that HiFi DSPs are already widely adopted current-generation TWS earbuds and Bluetooth headsets. He stressed that the new DSP is software compatible with the existing HiFi DSPs, which was important since the company has 160 partners on HiFi DSPs.
For next generation hearables, the advent of LC3 and wider market trends, plus AI-enabled speech and voice algorithms, will most likely drive expanding use cases for analytics and better sound quality in TWS earbuds. This is where the company believes its HiFi 1 DSP will enable the low energy consumption that will make always-on and always-listening capabilities to the mass market.
- Renesas and Syntiant offer joint voice-controlled vision AI solution
- How audio edge processors enable voice integration in IoT devices
- Design considerations for low-power, always-on voice command systems
- Add voice on a microcontroller without having to code
- Smarter voice-data processing yields better battery life
- How extensive signal processing chains make voice assistants ‘just work’
- Algorithms and hardware power rise of voice control
- Better audio processing at the edge