Cutting through the noise
Beyond merely isolating individual speakers for maximum clarity, the system goes a step further by eliminating unwanted sounds from the audio stream going out to remote participants.
Coral’s Edge TPU enables multi-channel noise cancellation technology called TrueVoice® —to let every voice come through loud and clear—while filtering out distracting sounds, such as snack wrappers, sirens, and typing.
The Edge TPU coprocessor allows the system to process audio locally as the system cannot rely on cloud processing. That’s because the large number of channels the system supports (up to 44, including channels provided by the mic pods) would create unacceptable lag and make conversations jumpy. “You can get similar performance at much higher cost and much higher power using a GPU, for example, but we are very power-limited,” Ribeiro explains.
Coral’s low power processing is ideally suited to meetings, Varghese says, for two reasons. First, the Coral chip’s low wattage doesn’t produce enough heat to require fans for cooling. No fans mean no extra noise to clutter up meeting audio. Second, low-power processing enables the new system’s minimalist, hassle-free setup. It runs on power over Ethernet, or PoE, reducing the number of cords and freeing up power outlets.
Both the Series One Smart Audio Bar and Meet Compute System feature the Edge TPU. In the Compute System, Coral’s new M.2 Accelerator with Dual Edge TPUs provides double the number of inferences per second as a single TPU, powering applications at a blistering eight trillion operations per second, using a total of only four watts of power. The new low-power powerhouse is great for vision processing applications and will enable an expanding array of capabilities for Series One Meet hardware.
Like the Chromebox, the new Smart Audio Bar also includes dual Edge TPU Accelerators soldered directly on board. Future-proofing is the name of the game here, as a single Edge TPU would do the job of isolating audio from individual meeting participants and filtering out noise. Including additional accelerators from the get-go will enable future capabilities, for example by allowing more than one machine learning model to run at the same time.