CLIP colloquium: Vikram Mitra (SRI International)

Speech processing applications such as speech recognition, keyword spotting, language recognition, speaker identification etc. are crucial for data analytics, voice operated systems, biometrics and information triage. Current speech processing systems perform well under matched/clean acoustic conditions but can degrade rapidly even with minor environmental changes. In particular, systems continue to be highly sensitive to acoustic variations due to noise, reverberation, different receiving-transmitting devices, multiple-speakers, etc. For real-world applications speech processing systems must perform reliably under greatly varying environmental conditions that are part of our everyday life. This talk will present SRI’s recent work on improving robustness of speech processing systems by exploiting findings in speech perception, machine learning and gestural phonology. We will talk about “speech-perception” motivated robust acoustic features developed in our lab that have demonstrated noise/channel robustness in multiple speech processing tasks. We will present a set of features (motivated by “gestural phonology”) that dynamically estimates vocal tract constriction location and degree in time, and talk about how they were used in some speech processing tasks. We will share details about some acoustic modeling approaches that we have found to be quite robust to background distortions. Finally, we will present results on some standard speech processing tasks and will investigate how the proposed systems perform under varying and unseen data conditions.