[an error occurred while processing this directive] [an error occurred while processing this directive]
Michael Mandel

Detailed models for understanding speech in noise

Michael Mandel, 1/28/2015, 4:00-5:00pm, BI 4369

The human ability to understand speech in noise far outstrips current automatic approaches, despite recent technological breakthroughs. This talk presents two projects that use detailed models of speech to begin to close this gap. The first project investigates the human ability to understand speech in noise using a new data-driven paradigm and is able to identify the specific spectro-temporal "glimpses" of individual speech utterances that are crucial to their intelligibility. By formulating intelligibility prediction as a classification problem, it is also able to successfully extrapolate these predictions to new recordings of the same and similar words. The second project aims to reconstruct damaged or obscured speech using a detailed prior model, a concatenative speech synthesizer. By learning a deep neural network affinity function between clean and noisy chunks of audio, the model can identify a coherent sequence of elements from a dictionary of clean speech that best match a noisy observation. The resulting noise-free syntheses have speech quality almost as high as the original clean speech, although with lower intelligibility. When combined, these projects will directly utilize knowledge and data from human listening studies to improve the noise robustness of machine listeners. [an error occurred while processing this directive]