Enhancement and Analysis of Conversational Speech: JSALT 2017

Automatic speech recognition is more and more widely and effectivelyused. Nevertheless, in some automatic speech analysis tasksthe state of the art is surprisingly poor. One of these is "diarization",the task of determining who spoke when. Diarization is key toprocessing meeting audio and clinical interviews, extended recordingssuch as police body cam or child language acquisition data, andany other speech data involving multiple speakers whose voices arenot cleanly separated into individual channels. Overlapping speech,environmental noise and suboptimal recording techniques make theproblem harder. During the JSALT Summer Workshop at CMU in2017, an international team of researchers worked on several aspectsof this problem, including calibration of the state of the art, detectionof overlaps, enhancement of noisy recordings, and classification ofshorter speech segments. This paper sketches the workshops results,and announces plans for a "Diarization Challenge" to encourage furtherprogress.