This Small Business Innovation Research (SBIR) Phase I research project focuses on a novel approach that uses multiple microphones to spatially separate speech and hence enhance speaker identification and speech recognition accuracy. The first component of the project will apply feature-driven adaptive beam forming to improve speech recognition accuracy in noisy and reverberant environments. Unlike conventional beam forming methods, this proposed method performs weight adaptation based on feedback from the feature extraction and recognition modules. The arrays can eliminate directional interferences and reduce reverberant effects. For the second component of the proposed project, a new continuously-variable mask estimation method will be used to estimate which parts of the separated speech contain speech with good signal-to noise ratio (SNR), applying speech reconstruction to repair these regions. The repaired speech will then be fed into the speech recognition/speaker identification engine. Thirdly, the project will apply ideas motivated by human binaural perception.
The potential market for robust speaker identification and speech recognition is projected to be large, ranging from voice dictation in noisy offices, voiceprint verification for financial transactions, biometrics, speech recognition in automated voice response systems, voice-based search for mobile devices, and several others. Additionally there are many military applications in military command, control, and communications segments, where it is important to convey the correct commands to the other recipients.
