Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

SPEECH UNDERSTANDING SYSTEMS (Report No. 7)

IP.com Disclosure Number: IPCOM000128804D
Original Publication Date: 1976-Dec-31
Included in the Prior Art Database: 2005-Sep-19
Document File: 11 page(s) / 37K

Publishing Venue

Software Patent Institute

Related People

W.Woods: AUTHOR [+9]

Abstract

Much of this past quarter was spent in designing, implementing, and testing the interface between the ^,^ Acoustic-Phonetic Experiment Facility (APEF) and the Acoustic-Phonetic Recognition (APR) program. The goal of this interface is to allow the APR to correctly adjust the scores for each phoneme against each segment in the segment lattice according to the particular acoustic feature values ^^~^ found within that segment. As a test of the effect of this individual adjustment to phoneme scores, we attempted to discriminate among the '""~" three nasal consonants. The APR program previously used conventional threshold decisions to choose one of several labels for each segment. The scores for particular phonemes , "" were determined by the statistics of the confusions between these segment labels and the correct phonemes.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 11% of the total text.

Page 1 of 11

THIS DOCUMENT IS AN APPROXIMATE REPRESENTATION OF THE ORIGINAL.

SPEECH UNDERSTANDING SYSTEMS

Quarterly Technical Progress Report No. 7

W.Woods L. Bates B. Bruce C. Cook J. Klovstad R. Schwartz J. Wolf

May 1976 to 31 July 1976

ARPA Order No. 2904 Contract No. N00014-75-C-0533

Program Code No. 5D30 Principal Investigator: William A. Woods (617) 491-1850 x361

Name of Contractor: Scientific Officer: Bolt Beranek and Newman Inc. Marvin Denicoff Effective Date of Contract: 'title: ' 30 October 1974 SPEECH UNDERSTANDING SYSTEMS Contract Expiration Date: QTPR Editor: 29 October 1976 Bonnie Nash-Webber (617) 491-1850 x227 Amount of Contract: $1,966,927

Sponsored by Advanced Research Projects Agency

ARPA Order No. 2904

This research was supported by the Advanced Research Projects Agency of the Department of Defense and was monitored by ONR under Contract No. N00014-75-C-0533.

A. Acoustic-Phonetic Recognition

Much of this past quarter was spent in designing, implementing, and testing the interface between the ^,^ Acoustic-Phonetic Experiment Facility (APEF) and the Acoustic-Phonetic Recognition (APR) program. The goal of this interface is to allow the APR to correctly adjust the scores for each phoneme against each segment in the segment lattice according to the particular acoustic feature values ^^~^ found within that segment. As a test of the effect of this individual adjustment to phoneme scores, we attempted to discriminate among the '""~" three nasal consonants. The APR program previously used conventional threshold decisions to choose one of several labels for each segment. The scores for particular phonemes , "" were determined by the statistics of the confusions between these segment labels and the correct phonemes. The highest

`"'" scoring nasal phoneme was correct: 70-75% of the time. With the non-parametric modeling procedure, which uses information from the APEF, the first choice nasal was

,.,." correct 90% of the time. What is more important is that for those segments which were correct, the scores on the other

~"'" nasals were often decreased very sharply. For those segments where the first choice was incorrect, the correct nasal had a score near the top scoring nasal.

"`.,' BBN Report No. 3359 Bolt Beranek and Newman Inc

Bolt, Beranek & Newman, Inc. Page 1 Dec 31, 1976

Page 2 of 11

SPEECH UNDERSTANDING SYSTEMS (Report No. 7)

Speaker Normalization

During the past quarter we also discussed several possible speaker/recording environment normalization procedures. Currently, the APR. is speaker independent. That is, any normalizing parameters used are derived

`""'"" directly from the utterance being recognized with no other knowledge about the speaker. While these techniques can perform quite well, the APR could be somewhat more accurate if there were some outside knowledge about this particular speaker. Some of the normalization procedures we discussed `"""" are as follows r-- 1) Using a carefully designed, phonetically balanced...