Canadian startup Picovoice recently launched a speech recognition engine that can run locally on small CPUs or microcontrollers in small devices like home appliances instead of using the cloud over the internet.
“Processing speech in the cloud has raised major privacy concerns with the uploading and handling of personal voice data," Picovoice said in a Sept. 17 blog. “Cloud speech recognition also has fundamental limitations with cost-effectiveness, latency and reliability.”
Picovoice has created software to run on commodity hardware as small as a $5 Raspberry Pi Zero which recognizes more than 200,000 words in real-time, the company said. It claimed its accuracy was a word error rate of about 10%, compared to about 8% for both Amazon and Mozilla, and better than Google speech-to-text at 12% and CMU PocketSphinx at 15%.
The company also claimed its software runs 23 times faster and consumers 53 times less memory than Mozilla DeepSpeech.
Picovoice noted that licensing its software could save more than 10 times the cost of cloud-based speech-to-text services that usually charge more than $1 per hour. Picovoice President Alireza Kenarsari-Anhari said the Picovoice approach will work well for small devices needing speech recognition. In an example he shared with EE Times, he said a public cloud service on a voice-activated coffeemaker would cost $15 a year, per appliance, if used 10 times a day. “You can do this for free if you use the resources you already have on the CPU on your coffeemaker,” he said. The Picovoice software uses language specific to each device which creates efficiency to the point where the model is less than half a megabyte and can be run on a sub-$1 MCU, he said.
Picovoice is earning revenue from LG, Whirlpool and Local Motors.
In April, Sensory introduced a speech recognition platform that works on the edge and doesn’t require a cloud connection.