From: Malcolm Slaney <malcolm@ieee.org>
Date: Sun, 15 Apr 2001 21:21:50 -0700
Subject: Understanding Speech at CCRMA Hearing Seminar
Message-Id: <5.0.0.25.2.20010415071235.00ac8508@worldserver.com>
This coming Thursday, Steve Greenberg kicks off the CCRMA Hearing Seminar
schedule with a talk about how man and machine perceive speech.
Understanding speech is certainly one of our most important perceptual
tasks. Yet we do not have enough information about how humans perceive
speech to form a complete model.
Speech perception by machines is even harder. While there are speech
recognition products in the market, they work well only in the best of
constrained situations. Perhaps their pattern recognition systems need
better features to describe the speech signal.
At this Thursday's Hearing Seminar, Steve will be describing evidence for
new dynamic features of speech.
Steve is a linguist, neurophysiologist and now speech researcher at the
International Computer Science Institute at UC Berkeley. He is an editor
of the book "Computational Models of Hearing" which will be published this
summer.
Who: Steven Greenberg (ICSI Berkeley)
What: The essential cues for understanding spoken language
When: Thursday April 19, 2001 at 11AM
Where: CCRMA Library, Top Floor of the Knoll at Stanford
I've also appended an invitation for the CCRMA Open House next month. See
you at CCRMA.
-- Malcolm
**** Steve Greenberg talking about speech perception at CCRMA ******
What are the essential cues for understanding spoken language?
Steven Greenberg
International Computer Science Institute,
1947 Center Street, Berkeley, CA 94704
Classical models of speech recognition (by both human and
machine) assume that a detailed, short-term analysis of the
acoustic signal is essential for accurately decoding spoken
language. Several lines of evidence call this assumption into
question :
(1) Intelligibility is relatively unimpaired when the
frequency spectrum is distorted under a wide range of
conditions (including cross-spectral asynchrony,
reverberation, waveform time reversal and selective
deletion of 80% of the spectrum),
(2) the acoustic properties of spontaneous speech
rarely conform to canonical patterns associated with specific
phonetic segments, and
(3) automatic-speech-recognition phonetic classifiers
often require ca. 250 ms of acoustic context
(spanning several segments) to function reliably.
This pattern of evidence suggests that the essential cues for understanding
spoken language are largely dynamic in nature, derived from
(1) the complex modulation spectrum (incorporating both
amplitude and phase) below 20 Hz,
(2) segmentation of the speech signal into syllabic intervals
between 50 and 400 ms, and
(3) a multi-time-scale, coarse-grained analysis of phonetic
constituents into features based on voicing, manner and place
of articulation.
-----------------------------------------------------------------------
The Center for Computer Research in Music and Acoustics (CCRMA) at
Stanford University is a multi-disciplinary facility where composers and
researchers work together using computer-based technology both as an
artistic medium and as a research tool. Graduate and undergraduate
students from the Music and Electrical Engineering Departments, as well as
visiting researchers and composers from around the world, conduct research
and compose electroacoustic music at CCRMA.
On May 10, 2001, CCRMA will hold a free Open House featuring lectures and
demonstrations of current research and compositional activities underway
at CCRMA. Research topics will include physical modeling sound synthesis,
source separation, musical instrument design, audio streaming over the
Internet, musical databases, and audio retrieval and classification.
The annual Open House is an opportunity for companies to get acquainted
with the work being done at CCRMA, and a chance to consider becoming an
Industrial Affiliate of CCRMA. In addition, the event allows CCRMA
students to interact with company representatives during the demo session
and at the catered barbeque.
CCRMA has been offering an Industrial Affiliates Program for the music and
audio industry since 1987. Industrial Affiliate benefits include a CD-ROM
compendium of materials from our annual research Open House, copies of all
research reports published by CCRMA, and early access to papers submitted
for publication, at the lowest membership level. For higher membership
levels, additional benefits include free participation in CCRMA Summer
Workshops, access to restricted portions of CCRMA's web site, and access
to CCRMA courses. There is also a membership level which involves direct
sponsorship of an engineering student, which allows a company
representative residency at CCRMA, and can include collaborative research.
If you would like to attend the 2001 CCRMA Open House, please visit the
CCRMA Affiliates web page, http://www-ccrma.stanford.edu/affiliates/, and
fill out the online RSVP form. If you would like more information or have
questions about the Open House or Industrial Affiliates Program, please
contact Charles Nichols at cnichols@ccrma.stanford.edu or by phone at
(650) 723-4971 ext. 354.