Industry consortium EEMBC introduces end-to-end audio benchmark for smart devices
Portland, United States –Â EEMBC has announced the release of AudioMark, the initial end-to-end benchmark specifically focused on audio processing.
AudioMark offers dramatic improvements in accuracy, portability and relevance over existing audio-related benchmarks, especially for critical tasks such as speech recognition. Developed in close collaboration with embedded processing providers Arm, Renesas, Infineon, Synopsys, onsemi, STMicroelectronics, Texas Instruments and Intel, AudioMark incorporates beamforming and direction-of-arrival algorithms, written by some of the world’s top DSP experts.
“As digital signal processing and machine learning capabilities in modern microcontrollers advance, system designers need a benchmark that can help evaluate the AI and DSP performance of these devices,” says Reinhard Keil, senior director of embedded technology, Arm. “EEMBC AudioMark is the result of close ecosystem collaboration, and addresses these requirements to provide system designers with realistic application scenarios, such as voice command control and smart speakers, as they develop innovative new products and solutions.”
Several features and capabilities set AudioMark apart from existing benchmarks. It utilises a model that can take both MCU and DSP processing into account, reflecting many companies’ increased reliance on DSP accelerators when designing for audio applications. It also uses a smaller 8-bit neural net, designed by Arm, which is ideal for memory- and power-constrained IoT edge applications. Reference code is provided right out of the box, but the key to AudioMark’s flexibility is the distinct porting layer that makes it easy to adapt the benchmark to a wide variety of platform hardware.
AudioMark also differs from most other EEMBC kernel-specific benchmarks in that it measures end-to-end performance on a variety of audio processing tasks, rather than focusing on a single function. “Audio processing performance depends on more than just the individual steps in the pipeline,” says EEMBC president Peter Torelli. “It was clearly necessary to develop an end-to-end benchmark for this critical task.” It is designed with several flexible parameters, to encompass variations such as single versus multiple Keyword Spotting (KWS), single versus multiple audio streams, and fixed- or floating-point computations
One of AudioMark’s most far-reaching updates, though, is in its scoring system. In addition to measuring function throughput speed for tasks like KWS, the benchmark also reports number of function calls per second relative to the processor clock speed. This “AudioMarks per Megahertz” score means that every embedded processor undergoing the AudioMark benchmark can be evaluated on both audio processing speed and processing efficiency. This is a critical metric for low-power, battery-operated devices, which form the core of the rapidly expanding IoT device industry.
As with all EEMBC benchmarks, involvement from members was key to making AudioMark relevant and useful. “EEMBC members are the world leaders in embedded processing,” says Torelli, “and their expertise and input has made AudioMark what it is: the new gold standard in benchmarking for audio processing.”
Comment on this article below or via Twitter @IoTGN