We're sorry – we can't find the page you're looking for.

We may have removed the page when we redesigned our website, the link you clicked might be old and does not work anymore, or you might have accidentally typed the wrong URL in the address bar.

Please visit Malaspina Labs Inc. Home Page. and follow the links from there.

Malaspina demonstrates 59% improvement in Automatic Speech Recognition for mobile phones

August 2012

Malaspina Labs has demonstrated the effectiveness of its VoiceBoost model based speech discrimination in improving the performance of Automatic Speech Recognition systems on mobile phones in noisy environments.

By isolating speech-of-interest from background noise (including background speech), VoiceBoost has been shown to improve word recognition rates of two leading Automatic Speech Recognition systems used in mobile phones. Operating at under 30 MIPS and 12ms latency on a Texas Instruments OMAP 4460 processor, VoiceBoost allowed the Google Voice Search ASR engine to correctly recognize 47% of words incorrectly recognized without the VoiceBoost processing (47% error recovery rate), and allowed Nuance's NDEV Mobile ASR to correctly recognize 59% of words incorrectly recognized without the VoiceBoost processing (59% error recovery rate). Maximum benefit was seen between 10dB and 15dB SNR.

Testing was performed using the National Institute of Standards and Technology's Speech Recognition Scoring Toolkit (SCTK) Version 2.4.0 using test files consisting of in-situ recorded traffic and babble noise ranging from 0db to 25dB SNR.

Earlier this year VoiceBoost was also shown to improve speech quality by 36% over existing 3G phone and network noise reduction in real-world high noise environments.

VoiceBoost is effective with single or dual microphone designs and both near-field (phone at mouth) and arms-length (hands-free) speaking, without the need for additional voice processing hardware.