integrated processing unit, particularly for connected speech recognition systems

2
In the disclosed method, the location and amplitude of the k th pulse are dependent onlyon previously foundpulses whichlie within a constrained intervalof the k th pulse, all within the currentfundamental period.The patent offers a more efficient method of computing thepulse parameters.-- DLR 4,907,279 43.72.Ja PITCH FREQUENCY GENERATION SYSTEM IN A SPEECH SYNTHESIS SYSTEM Norio Higuchi, Seiichi Yamamoto, and Toru Shimizu, assignors to Kokusai Denshin Denwa Company 6 March 1990 (Class381/52); filed in Japan31 July 1987 This system computes fundamental frequency profiles for use in a speech synthesis system based onphrase andaccent marks entered together with the text to be synthesized. The method is based on that described by Fujisaki for the Japanese language, but offers a more efficient methodof (b) (c) (d) (e) lO• lD2 To. To•, t A B• B(=B1 '*'B2 ) A T• T2• T•2 T22 i• I I i ,•41 i ' i i ; ' Ai+A2•' B1 +B2 i t computing the acoustic effects of the phrase and accentmarkers.The im- proved efficiency seems to stem largely froma liberal use of lookup tables.-- DLR 4,926,484 43.72.Kb CIRCUIT FOR DETERMINING THAT AN AUDIO SIGNAL IS EITHER SPEECH OR NON- SPEECH Yoshitomo Nakano, assignor to SonyCorporation 15 May 1990 (Class 381/56); filed in Japan 13 November1987 Described hereis a method of discriminating speech from nonspeech signals foruse inavoice-activated recording device. Likemany other speech detector methods, this one is based ona comparison oftheshort-term signal energy witha threshold value. In this patent, thethreshold level isadjusted so as to control theoverall percentage oftimethattheinput level exceeds the threshold. If theinput isbelow threshold for more than 3 sof the first30 sof recording, thenthethreshold levelislowered. As a result, the threshold will tend to trackjust above the levelof a steady ambient noise.--DLR 4,982,341 43.72. Kb METHOD AND DEVICE FOR THE DETECTION OF VOCAL SIGNALS Pierre A. Laurent, assignor to Thomson CSF 1 January 1991 (Class364/517); filed in France 4 May 1988 This methodfor detecting the presence of a speech signal locates the maxima within each 20-ms frameof the pre-emphasized inputandof a high- pass-filtered versionof that pre-emphasized speech. Both the high-pass maximumand its ratio to the pre-emphasis maximumare allowedto in- crease freely without constraint,but are low-pass filtered if decreasing. These and other comparable filtering steps provide several long-term threshold values that are used to decide whether the input isin fact a speech signal.--DLR 4,905,286 43.72.Ne NOISE COMPENSATION IN SPEECH RECOGNITION Nigel C. Sedgewick and John N. Holmes, assignors to National Re- search DevelopmentCorporation 27 February 1990 (Class381/43); filed in the United Kingdom 4 April 1986 This technique, for improving the performance of a Markov speech recognizer with noisy input,uses analternate method of computing spectral framedistances, depending on whether the current inputin each frequency band is above or below a noise threshold value. The basic distance measure is a probability density function or likelihood of seeing the currentinput value given each of several possible candidate acoustic states. If the current input is belowthe noise level, the likelihoodalgorithmis alteredto reflect the lack of knowledge of that frequency band's actualsignal level.--DLR 4,905,288 43.72.Ne METHOD OF DATA REDUCTION IN A SPEECH RECOGNITION Ira A. Gerson and Brett L. Lindsley, assignors to Motorola, Incorpo- rated 27 February 1990 (Class381/43); filed 3 January 1986 This improvement for isolated-word speech recognition systems at- tempts to reduce the datarateof incoming speech databy collapsing certain input frames soasto reduce the redundancy. A sequence of representative substitute frames replaces a longer sequence of frames occurring initially in ; ; FEATURE ' J FRAMES 510 • CLUSTER 51//' D1 • \/,,.,.•/05 AVERAGE [ \ the utterance. As the substitute frames are computed, a distortion measure is associated with each replacement frame, which must remain below a distortion threshold value.--DLR 4,907,278 43.72.Ne INTEGRATED PROCESSING UNIT, PARTICULARLY FOR CONNECTED SPEECH RECOGNITION SYSTEMS Riccardo Cecinati et aL, assignors to Dello Stato Italiano 6 March 1990 (Class 381/43); filed in Italy 25 June 1987 This patent presents the logic design for an integrated circuit chip applicable to a connected speech recognizer usinghierarchical dynamic 1199 J. Acoust.Soc. Am., Vol. 92, No. 2, Pt. 1, August1992 Patent Reviews 1199 Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 141.212.109.170 On: Tue, 16 Dec 2014 13:33:00

Upload: riccardo

Post on 11-Apr-2017

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Integrated processing unit, particularly for connected speech recognition systems

In the disclosed method, the location and amplitude of the k th pulse are dependent only on previously found pulses which lie within a constrained interval of the k th pulse, all within the current fundamental period. The patent offers a more efficient method of computing the pulse parameters.-- DLR

4,907,279

43.72.Ja PITCH FREQUENCY GENERATION SYSTEM IN A SPEECH SYNTHESIS SYSTEM

Norio Higuchi, Seiichi Yamamoto, and Toru Shimizu, assignors to Kokusai Denshin Denwa Company

6 March 1990 (Class 381/52); filed in Japan 31 July 1987

This system computes fundamental frequency profiles for use in a speech synthesis system based on phrase and accent marks entered together with the text to be synthesized. The method is based on that described by Fujisaki for the Japanese language, but offers a more efficient method of

(b)

(c)

(d)

(e)

l O• l D2 To. To•, t

A B•

B(=B1 '*'B2 )

A

T• T2• T•2 T22

i• I I i ,•41 i

' i i

;

' Ai+A2•' B1 +B2 i

t

computing the acoustic effects of the phrase and accent markers. The im- proved efficiency seems to stem largely from a liberal use of lookup tables.-- DLR

4,926,484

43.72.Kb CIRCUIT FOR DETERMINING THAT AN

AUDIO SIGNAL IS EITHER SPEECH OR NON- SPEECH

Yoshitomo Nakano, assignor to Sony Corporation 15 May 1990 (Class 381/56); filed in Japan 13 November 1987

Described here is a method of discriminating speech from nonspeech signals for use in a voice-activated recording device. Like many other speech detector methods, this one is based on a comparison of the short-term signal energy with a threshold value. In this patent, the threshold level is adjusted so as to control the overall percentage of time that the input level exceeds the threshold. If the input is below threshold for more than 3 s of the first 30 s of recording, then the threshold level is lowered. As a result, the threshold will tend to track just above the level of a steady ambient noise.--DLR

4,982,341

43.72. Kb METHOD AND DEVICE FOR THE DETECTION OF VOCAL SIGNALS

Pierre A. Laurent, assignor to Thomson CSF 1 January 1991 (Class 364/517); filed in France 4 May 1988

This method for detecting the presence of a speech signal locates the maxima within each 20-ms frame of the pre-emphasized input and of a high- pass-filtered version of that pre-emphasized speech. Both the high-pass maximum and its ratio to the pre-emphasis maximum are allowed to in- crease freely without constraint, but are low-pass filtered if decreasing. These and other comparable filtering steps provide several long-term threshold values that are used to decide whether the input is in fact a speech signal.--DLR

4,905,286

43.72.Ne NOISE COMPENSATION IN SPEECH

RECOGNITION

Nigel C. Sedgewick and John N. Holmes, assignors to National Re- search Development Corporation

27 February 1990 (Class 381/43); filed in the United Kingdom 4 April 1986

This technique, for improving the performance of a Markov speech recognizer with noisy input, uses an alternate method of computing spectral frame distances, depending on whether the current input in each frequency band is above or below a noise threshold value. The basic distance measure

is a probability density function or likelihood of seeing the current input value given each of several possible candidate acoustic states. If the current input is below the noise level, the likelihood algorithm is altered to reflect the lack of knowledge of that frequency band's actual signal level.--DLR

4,905,288

43.72.Ne METHOD OF DATA REDUCTION IN A SPEECH RECOGNITION

Ira A. Gerson and Brett L. Lindsley, assignors to Motorola, Incorpo- rated

27 February 1990 (Class 381/43); filed 3 January 1986

This improvement for isolated-word speech recognition systems at- tempts to reduce the data rate of incoming speech data by collapsing certain input frames so as to reduce the redundancy. A sequence of representative substitute frames replaces a longer sequence of frames occurring initially in

; ; FEATURE ' J FRAMES

510 • CLUSTER 51//'

D1 • \/,,.,.•/05 AVERAGE [ \

the utterance. As the substitute frames are computed, a distortion measure is associated with each replacement frame, which must remain below a distortion threshold value.--DLR

4,907,278

43.72.Ne INTEGRATED PROCESSING UNIT, PARTICULARLY FOR CONNECTED SPEECH

RECOGNITION SYSTEMS

Riccardo Cecinati et aL, assignors to Dello Stato Italiano 6 March 1990 (Class 381/43); filed in Italy 25 June 1987

This patent presents the logic design for an integrated circuit chip applicable to a connected speech recognizer using hierarchical dynamic

1199 J. Acoust. Soc. Am., Vol. 92, No. 2, Pt. 1, August 1992 Patent Reviews 1199

Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 141.212.109.170 On: Tue, 16 Dec 2014 13:33:00

Page 2: Integrated processing unit, particularly for connected speech recognition systems

programming. This form of recognizer involves a large number of matrix operations at the lowest level. According to the patent, the recognition task can be partitioned so that word evaluations are performed by multiple slave math units, while a master processor computes the best-matching path through the syntactic structure.mDLR

describes a curved hand-held bow 10 with a layer 15 of friction material (e.g., natural or synthetic horsehair strands) for stroking selected string 19 in a planar array of strings 18. This converts a guitar into a bowed string instrument.mDWM

5,092,215

43.75.De GUITAR BOW

William Mello, Bedford, MA, and John Hendrickson, Easton, MA 3 March 1992 (Class 84/325); filed 26 December 1990

Bowed string instruments generally have the cross section of their strings arranged in an arc to facilitate the use of a straight bow. Plucked string instruments generally have the cross section of their strings in a plane since they usually have straight frets for fingering purposes. This patent

20

.

IO

t$

\/ I•r 19 18

5,090,291

43.75.Tv MUSIC SIGNAL TIME REVERSE EFFECT APPARATUS

Louis A. Schwartz, Ansonia, CT 25 February 1992 (Class 84/603); filed 30 April 1990

The envelope of a percussive musical tone has a rapid rise time and a slower decay time, depending upon the decay rate of the resonant tone source. A novel tonal effect can be produced by reversing the tone envelope. In earlier patents (4,003,285 and 4,160,402) to the same inventor, a digital electronic circuit applied prestored reversed envelope information to a mu- sical tone envelope in an attempt to produce the tonal effect of the reversed tone of a piano or guitar, for example. In the present patent provision is made for changes in the rate of decay during the duration of a single tone, as described for piano tone in the reviewer's articles in the Journal [ J. Acoust. Soc. Am. 19, 535 (1947) and 64, 1303 (1978) ]. The result can be only a first approximation to the actual reversed percussive tone since percussion has a cause and an effect sequence in which the timbre of the cause usually differs greatly from that of the effect.•DWM

1200 J. Acoust. Soc. Am., Vol. 92, No. 2, Pt. 1, August 1992 Patent Reviews 1200

Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 141.212.109.170 On: Tue, 16 Dec 2014 13:33:00