Knowles, the world's biggest supplier of microphones, said that it would start selling a voice processor that could let people talk to smart speakers during a noisy cocktail party, switch on lights with a few voice commands, or give noise suppression powers to headphones.
Knowles’ system-on-chip is aimed at curing the unnaturalness of voice controls in everything from speakers and alarm clocks to kitchen appliances and smartphones. The company could have potential customers in Amazon, Google, and Apple selling smart speakers imbued with digital assistants like Alexa, Home, and Siri.
The new product – unceremoniously called IA8508 – can pre-process human speech used by voice controls. Knowles snuck custom circuitry in the four-core processor to handle machine learning algorithms that would typically run on the cloud to sift through background noise, echoes, and reverb while amplifying and sharpening the speaker’s voice.
Knowles is trying to plant audio algorithms “closer to the human, and that means closer to the microphone,” said Mike Polachek, Knowles’ president of intelligent audio. “If you can move more computing from the cloud to the device and then from deeper in the device to the edge, you can better understand context and how many people are in the room.”
Three of Knowles’ four cores are digital signal processors. The high-performance core can be dialed to boost sensitivity to faraway voices, while the low-latency core can improve active noise cancellation. A low-power core can be programmed to listen to multiple keywords simultaneously – like “bake” and “preheat” in an oven – without smothering battery life. The last core is an Arm microcontroller.
In an interview, Polachek suggested that the chip enabled devices to process raw audio locally rather than in the cloud, resulting in lower latency and greater privacy. That is particularly useful, for example, in smart headphones that need to amplify faint sounds and suppress loud chatter in an office or on a busy city street.
Pairing voice control chips with its microphones could be lucrative for Knowles, which holds more than 40% of the microphone market. In that spirit, the Itasca, Illinois-based firm built its new processor with 5.7 megabytes of shared memory and ports for eight microphones to provide more contextual awareness to what its microphones hear.
Last year, sales of smart speakers in the fourth quarter were around 4.2 million, up from 600,000 in the previous fourth quarter, according to Strategy Analytics. To prepare for that market and others like it, Knowles paid $85 million in 2015 for Audience, which created a chip to listen for simple commands, detect glass breaking, or an infant’s crying.
Knowles competes with Cirrus Logic, which sells voice chips that nullify background noise and deduce the direction of sounds using four-microphone arrays. Synaptics’ Conexant unit does the same, and its silicon works with both Baidu and Amazon’s platform. Last month, Xmos raised $15 million to advance its multicore microcontrollers for voice recognition.
These chipmakers are trying to lower the computing cost of voice recognition for potential customers. Last month, at the Arm Research Summit, Google’s Pete Warden called for voice recognition chips that could grasp basic commands like on and off – and simple sounds like a grasshopper’s chirping or knocking in an industrial machine – and are inexpensive enough to be disposable.
Moreover, they are trying to exorcise technological demons like voice recognition over longer distances. “Five years ago, there was an assumption that you would be talking to a device within five feet, and many times within one or two feet,” said Todd Mozer, chief executive of Sensory, in an interview with Electronic Design last month.
“Over the last few years, people have said that they want to recognize voices from thirty feet away, which means more reverberations, more background noise, and higher signal-to-noise ratio,” Mozer said. Sensory sells keyword detection software for tiny microprocessors embedded in light bulbs, kitchen appliances, and other devices.
Sensory and other software suppliers are also Knowles' rivals. Typically, their code runs on Arm processors not DSPs, which are notoriously difficult to program. For that reason, Knowles has partnered with eight software firms – including Sensory and automotive start-up DSP Concepts – to port code to its new chip.
“Lots of these software algorithm companies see DSP companies as competition,” said Polachek. “But that software is already out there, and we wanted to create an environment where all that software can be reused and used on our hardware.” The IA8508 will enter full production in the first quarter next year.