Voice and AI Explosion Rocks CES

　　Voice, connectivity and AI took center stage at the Consumer Electronics Show last week. If this year’s CES is any indication, these three building blocks will compose the holy trinity of consumer electronics devices that will drive the market in 2018 and further into the future.

　　Voice assistants are now poised to move into wearables, headphones, baby monitors, lamps, TV remotes and vehicles. Paul Beckmann, founder and chief technology officer of DSP Concepts, told EE Times, “We are witnessing a Cambrian explosion around voice.”

　　At CES, Baidu, known as “China’s Google,” shouted out most loudly for voice by unveiling and opening to developers its Duer OS-based platform. Neither its voice-enabled lamp, ceiling-mounted projector nor screen need Alexa or Google Assist. A growing number of vendors are gravitating toward voice, as Baidu loves to say, at “China speed.”

　　Connectivity in consumer devices is already a given. The next necessity is the ability to “mix and match” different wireless networks, stressed Silicon Labs CEO Tyson Tuttle. Casually slapping onto IoT devices a connectivity chip originally designed for smartphones will no longer suffice, he explained. Systems need dynamic multi-protocol software and the ability to time-slice different wireless networks.

　　While AI dominates attention as a key enabler for highly automated vehicles, Gideon Wertheizer, CEO of Ceva, told us, “I see AI getting out of the fantasy world.” Vendors are now trying to “set up parameters to use AI to solve specific problems in a random environment,” he explained.

　　In other words, companies are learning to use AI in bite-size and apply it to specific tasks, rather than depending on AI to solve the world’s problems.

　　Voice goes on the road

　　Voice is going to be critical both in the home and on the road.

　　Bosch announced at CES its plan to seat its voice assistant behind the wheel. “We are putting an end to the button chaos in the cockpit,” declared Dirk Hoheisel, a member of the board of management of Robert Bosch GmbH. Elektrobit promised at CES that it will be among the first Amazon Alexa automotive software integrators.

　　As it unveiled its Duer OS-based Apollo 2.0 platform, dubbed “Android for automated vehicles,” Baidu asserted that voice assistance will be an integral part of the platform. Qi Lu, Baidu's vice chairman, said, “There will be no border between a home and a vehicle. Whatever you can do at home, you should be able to do it in cars.”

　　Whether at home or in vehicles, isolating voice and sending clear signals all the way to the cloud is very hard, Ceva’s Wertheizer pointed out. “We are surrounded by noise.” Naturally, solutions for homes and cars must be able to handle a set of very different noise environments.

　　No standard voice algorithms

　　Complicating matters is the absence of any standard voice algorithms in the industry. “Everyone has its own proprietary algorithms to deal with voice,” observed Ceva CEO Gideon Wertheizer.

　　System companies are going back to technical papers published in academia, scrambling to figure out how best to isolate voice. They need to optimize their algorithms to varying settings as they use different microphones and types of speakers.

　　Over at Ceva, Wertheizer said, “We had to build an atomic shelter-like studio” to study all the options and develop algorithms for beamforming, far-field and near field, acoustic echo cancelation and ambient noise reductions. System companies need “support, DSP and software” that can navigate a jungle where no standards exist, he said.

　　DSP Concepts’ CTO Beckman echoed the sentiment. As much as people love voice as a natural user interface, he said, “Unfortunately, it is one of the most challenging technologies for product designers to effectively implement.” He has already seen too many voice projects go horribly wrong, eventually ending up back on his drawing board.

　　Beckman, who cut his teeth as a research engineer at Bose Corp. for nine years, built his business as a consultant in the early 2000s. As he worked with clients on voice projects, he realized they need “a complete software solution” that performs very well, and “underlying technology that will allow them to differentiate.” But most critical was to give them the ability to tune up their systems, he noted.

　　As the voice market exploded, so did Beckman’s business. DSP Concepts is no longer a consultancy. The company now offers a complete set of algorithms as software libraries and debugging tools that can help them tweak their individual systems around the edges. “We offer tuning, integration and validation,” said Beckman.

　　With the company’s voice UI technology called Audio Weaver, DSP Concepts is the first third-party software company qualified by Amazon for Alexa products. Chin Beckmann, co-founder and CEO of DSP Concepts, told EE Times that an Audio Weaver-enabled voice assistant product has demonstrated — using only two microphones instead of seven used in Amazon’s Echo — that it can “hear” voice much more clearly than either Echo or Google Home.

　　Getting pragmatic about AI

　　Isolating voice is step number one, Wertheizer said, but other steps follow. A voice assistant must recognize the voice’s location and must be able to track it. Moreover, it needs to detect — and identify — who is speaking in the room.

　　Until recently, the cloud was assumed to be where all that processing and learning take place. That assumption will change in 2018.

　　Wertheizer explained, “I see people are becoming more pragmatic about AI. They want to do it on the edge” rather than in the cloud, in order to avoid such issues as privacy, latency and cost.

　　David Ku, MediaTek’s CFO, agreed. 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.” He told us at the show that the voice-assistant race already focuses on adding “intelligence” locally, to separate human from non-human voices, cancel music in the background, and recognize vocal patterns.

　　Ceva’s CEO said, “Consider a product from Petcube” — a company that designed an interactive Wi-Fi pet camera. It can monitor, talk to, and interact with a dog or cat through two-way audio and a 1080p HD video camera while the owner is not home. “I’m not sure if Petcube realizes that it’s an IoT company,” said Wertheizer. But clearly, in a connected product like this, the voice recognition system must be able to recognize a dog’s barking, and identify if the dog is under stress or in a crisis, he explained. In other words, the system needs smarts to learn.

　　Meet Neupro

　　While Ceva offers voice algorithms called ClearVox to designers of voice-enabled systems, it also knows this is only half of what system vendors want. System manufacturers want to integrate inside their IoT devices the ability to learn and do inferences, so that their products can continue to get smarter.

　　The market craves AI processors. To meet the demand, Ceva launched, at CES, NuePro, “a dedicated low-power AI processor family for deep learning at the edge." NuePro is a self-contained, specialized AI processors that scales in performance for a broad range of markets including IoT, smartphones, surveillance, automotive, robotics, medical and industrial.

　　Notably, Ceva is no novice in deep learning. NeuPro reportedly builds on Ceva’s experience in deep neural networks for computer-vision applications.

　　Wertheizer said the NeuPro AI processor is the first “non-DSP” technology Ceva has developed from the ground up. In announcing Neupro, “I was a little nervous,” he said. “But you need to understand that AI is not a signal processing problem,.

　　The NeuPro processor comes with two pieces of hardware — a NeuPro engine and NeuPro VPU (vector processing unit).  While the engine handles well-defined AI algorithms such as CNN, activation and normalization layers, the NeuPro VPU, a programmable vector engine, is an extension that runs proprietary AI algorithms — or algorithms that have not been invented yet — Wertheizer noted. “Rather than using GPU or CPU, we opted for this hardwired implementation, so that we can increase the AI processor’s utilization.”

　　Ceva claims that this new family of dedicated AI processors offers “a considerable step-up in performance, ranging from 2 Tera Ops per second (TOPS) for the entry-level processor and 12.5 TOPS for the most advanced configuration.”

　　Ceva said NeuPro AI processors will become available for licensing to its lead customers in the second quarter of 2018. The company plans general release in the third quarter.

　　Similarly, Taiwan’s MediaTek is getting ready to push AI on the edge with a new AI processor developed by a 2016 Taiwan startup called Intelligo, a MediaTek spinoff.

　　Billed as “an intelligent DNN voice processor,” the scope of the AI SoC designed by Intelligo is much more limited. The processor offers “configurable deep neural networks and highly efficient inference engine (1 TOPS per second per watt)," according to the company.

　　David Ku, MediaTek’s chief financial officer, said that his company is looking for a modest AI accelerator designed to recognize only 20 to 30 key words. MediaTek is promoting the idea of “decentralized processing” by installing voice and AI not just in a smart speaker like Echo or Google Home, but in a range of small devices — including light switches.