Compared to previous releases, Version 4.0 of TrulyHandsfree speech-recognition software from Sensory boasts an 80% increase in word accuracy. The embedded small-footprint voice-interface platform employs filter-bank features that improve front-end speech-feature extraction, along with new phrase-spotting techniques that ensure the core part of the user’s spoken request can be recognized in the middle of speech.
TrulyHandsfree V4.0 contains a new neural net engine that supports deep learning acoustic models, dramatically improving speech-recognition accuracy amidst real-world noise. The technology is able to fully operate with less than a 2-mA draw on the battery and as low as 1 mA with Sensory’s low-power sound-detection IP. Enhanced architectural scalability allows for low-power DSP implementations with secondary accuracy improvements at the operating system level.