CEVA WhisPro is a Low Power Wake Word Detection Solution at the IoT Chip Level
CEVA yesterday announced the availability WhisPro, a low-power wake word detection solution that resides at the chip level for IoT devices. The company describes itself as a leading licensor of intellectual property (IP) for digital signal processing platforms and artificial intelligence processors. CEVA technology is used in more than one billion chips annually produced by companies such as Intel and Samsung.
IoT devices are a particular focus for the company and WhisPro is a software package for AI processors that runs a low-power recurrent neural network locally for wake word detection. In short, it is designed to enable high fidelity wake word detection for devices that offer always-listening voice interaction, but wait to activate until the trigger phrase is spoken by a user. This is accomplished on-device without the need for internet connectivity. The product that is shipping now supports English and Mandarin Chinese is expected later this year. It is shipped with a software development kit (SDK).
Focus on Recognition Rate and Low Power Consumption
The solution is marketed as having no more than one false positive per hour and false negatives of only 5%. CEVA’s Youval Nachum, senior product marketing manager for audio and voice, said in an interview, “The false [negative] ratio is what I call the frustration ratio. It is when you ask for the voice assistant but it doesn’t wake up.” The 95% recognition rate is achieved “without the need for cloud connectivity.” Minimization of false positive addresses another issue. Each “wake-up” draws power. As a result, minimizing false positives helps reduce the wake word engine’s power consumption. This is an important consideration for many IoT devices.
Nachum also said WhisPro enables device makers the opportunity to offer multiple wake words to be operational simultaneously. This has been a challenge that Sonos has been unable to solve to date. The company said in late 2017 that its Sonos One smart speaker which initially launched with Alexa would support both Alexa and Google Assistant simultaneously on the device in 2018. A year later, the company acknowledged that it still wasn’t able to ship a product with Google Assistant and Alexa on the same device.
Our advantage is using a single engine to recognize multiple triggers. So, it reduces footprint.
WhisPro wouldn’t necessarily address Sonos’ problem in particular but is designed to support a similar user experience of multiple wake words which route the event to different outcomes or services. Nahcum commented, “Our advantage is using a single engine to recognize multiple triggers. So, it reduces footprint.” WhisPro also offers the ability to “work with any set of customer defined keywords” which means device makers can customize their wake word or trigger phrases.
Combining Wake Word and Noise Reduction
![](https://voicebot.ai/wp-content/uploads/2019/01/clearvox-plus-whispro.png)
Moshe Sheier, a CEVA marketing vice president, also pointed out that WhisPro was designed to work with the company’s existing ClearVox “front end noise reduction and acoustic echo cancelation software solution.” This enables pre-processing of speech before the wake word detection is activated. The noise reduction and acoustic echo cancelation facilitate higher fidelity for “barge in” where the device is delivering audio content and the user seeks to interrupt. Company materials say the combination of ClearVox and WhisPro deliver both performance and power consumption benefits.
Both Sheier and Nachum acknowledged that CEVA is entering an increasingly competitive field for on-chip solutions related to voice capabilities. However, they pointed out that data from Yole Development shows the demand for sound processor units is growing quickly due to the rise in smart home systems and their technology combined with relationships with chip makers positions them for success.
Follow @bretkinsella Follow @voicebotai
CES 2019 – Eight Interviews with Voice Innovators – Voicebot Podcast Ep 78