EEMBC is pleased to announce the development of AudioMark, the first end-to-end benchmark specifically focused on audio processing. Slated for release in 2023, AudioMark will be invaluable for developers of embedded microprocessors that cater to voice-controlled appliances, speech recognition, and other related audio processing tasks.

Voice interaction has grown from a specialized interface associated with intelligent speakers to something we encounter every day. As this demand for ubiquitous audio-processing has increased, the microprocessor industry has shifted to provide the technology that powers it

“AudioMark is a departure from EEMBC's kernel-specific benchmarks, for good reason: 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.”

Because audio processing encompasses so many different scenarios, AudioMark will be designed with several flexible parameters. A device that listens for a single word or phrase, for instance, has different processing needs than one that understands a multi-word vocabulary, and is more likely to process locally than send audio to the cloud. Other variables include single versus multiple audio streams, fixed- or floating-point computations, and if a DSP is used alongside the processor.

“Audio is a rapidly growing part of our business, and we need a reliable way of comparing our products with each other, and with our competitors',” says Dr. Ashutosh Pandey, Lead Principal Engineer at Infineon Technologies and the technical lead of the AudioMark working group at EEMBC. “But audio processing exercises so many aspects of compute, and can be pursued in many different ways. It's critical that the industry have an audio benchmark that's flexible, but coherent enough to be meaningful.”

Fortunately, many aspects of the audio processing pipeline are fairly consistent. Nearly all start with a microphone with 16-bit output at 44.1 kHz, perform spectral analysis via Fourier transform, and must apply echo and noise cancellation. Similarly, other processes such as beam-forming and direction-of-arrival, are also key pipeline components. The final steps may employ a local feature extractor and neural net, or be synthesized for delivery to a cloud-based endpoint.

In order to ensure that AudioMark strikes the right balance between flexibility and consistency, numerous expert advisors have been recruited from the EEMBC community. Intel, Arm, onsemi, Renesas, Infineon, STMicroelectronics, Synopsys, and Texas Instruments have already signed on to assist in AudioMark's development and testing, and their help will be invaluable.

With AudioMark still in early development, there's a huge opportunity for microprocessor makers to get involved in shaping the benchmark that will help shape the industry. “By providing information about their specific tasks and processes, EEMBC members will ensure that AudioMark is as relevant and useful as possible,” says Torelli, “making this an ideal time to join as a member—especially if audio processing is part of your business.”
 

— WebWireID290674 —