Ku acknowledged that voice assistants such as Amazon’s Echo and Google Home are “a hot idea.” While a mouse became the user interface of choice in the PC era, “touch” proliferated among smartphones as the preferred UI. Now fingers are giving way to vocal chords, he explained.
According to Ku, it’s a no-brainer for MediaTek to move into smart speakers like Echo or Google Home. “After all, those devices are a tablet without a screen,” he noted.
But MediaTek isn’t just satisfied with selling tablet chips for voice-assistant speakers like those made by Sony.
In contrast to Amazon’s push for cloud-to-cloud services in its Echo devices, MediaTek sees possibilities in a hybrid model of “edge and cloud.” The race already focuses on adding more “intelligence” locally, so that smart speakers can separate human voices from non-humans, cancel music in the background, and recognize vocal patterns.
Some of those building blocks for voice AIs are getting designed into smartphones, headphones and automotive platforms, he noted.
MediaTek’s tablet processor is designed into Sony’s Google Assistant.
DSP combined with neural network
MediaTek, however, wants to push the idea of AI integration further onto devices as small as light switches.
Reportedly, the foremost consumer use of smart speakers is telling light switches to turn on and off. So, Ku asked, “Why not add that small AI capability to light switches?”
When connected lights were introduced, consumers were initially happy to do a little home demo, showing off how they can control the lights via smartphones. After a few days, though, Ku said, “What’s the fun in doing that? You might as well physically flip a switch.”
But if you can talk directly to a light switch, it’s a different story, he said. Without installing a voice assistant system in every room, “you can just say to the switch to turn itself off.”
Required for such a switch will be a small DSP and neural network that can recognize only 20 to 30 key words, he said. The DSP functions as a deep learning accelerator. The idea of a “decentralized processing unit” is important to MediaTek opening the door to sell chips into more than a billion devices.
Enabling AI on the edge is an AI processor and algorithms designed by a Taiwan startup called Intelligo. The company’s 30-member team spun off from MediaTek in 2016.
Intelligo designed an SoC called Debussy that functions as an intelligent DNN voice processor. The core of Debussy is Intelligo’s AI processor, iGo.
Obviously, installing AI in such a small device like a light switch comes with constraints that limit its scope, function, cost and power consumption. Ku acknowledged, “We are still in an early stage of AI. There’s still lots to learn.”
According to Intelligo, iGo offers “configurable deep neural networks and highly efficient inference engine (1 TOPS per second per watt).”
Intelligo also has a suite of iGo voice software. Included are: bi-directional voice quality enhancement, stationary and non-stationary noise cancellation, far-field processing, keyword spotting, voice verification and wake-on-voice.
The SoC provides a host of audio processing functions including acoustic echo cancellation, dynamic range control, equalization and sample rate conversion.
Asked about the team at Intelligo, Ku said, “We thought it was best to be spun out of MediaTek, because the startup environment would allow them to move much faster, thus accelerating their development.” MediaTek is an investor in Intelligo. Intelligo today already has customers other than MediaTek.
— Junko Yoshida, Chief International Correspondent, EE Times