US 8,942,977 B2Granted

System and method for speech recognition using pitch-synchronous spectral parameters

OwnerChengjun Julian Cheninventor

Life of this patent

Estimating the term…

Loading patent…

Abstract

The present invention defines a pitch-synchronous parametrical representation of speech signals as the basis of speech recognition, and discloses methods of generating the said pitch-synchronous parametrical representation from speech signals. The speech signal is first going through a pitch-marks picking program to identify the pitch periods. The speech signal is then segmented into pitch-synchronous frames. An ends-matching program equalizes the values at the two ends of the waveform in each frame. Using Fourier analysis, the speech signal in each frame is converted into a pitch-synchronous amplitude spectrum. Using Laguerre functions, the said amplitude spectrum is converted into a unit vector, referred to as the timbre vector. By using a database of correlated phonemes and timbre vectors, the most likely phoneme sequence of an input speech signal can be decoded in the acoustic stage of a speech recognition system.

Claims12

Dependent claims

A method of automatic speech recognition to convert speech signal into text using one or more processors comprising:

A) segmenting the speech signal into pitch-synchronous frames, wherein for voiced sections each said frame is a single pitch period;
B) for each frame, equalizing the two ends of the waveform using an ends-matching program;
C) generating an amplitude spectrum of each said frame using Fourier analysis;
D) transforming each said amplitude spectrum into a timbre vector using Laguerre functions;
E) performing acoustic decoding to find a list of most likely phonemes or sub-phoneme units for each said timbre vector by comparing with a timbre vector database;
F) decoding the sequence of the list of the most likely phonemes or sub-phoneme units using a language-model database to find out the most likely text;
wherein the segmenting of the speech-signal is based on an analysis of the speech signals using an asymmetric window which includes:
a) conducting, for a speaker, a test to find the best size of the asymmetric window;
b) convoluting the speech-signal with the said asymmetric window to form a profile function;
c) picking up the maxima in the said profile function as segmentation points;
d) extending the segmentation points to unvoiced sections.

8 more elements

A system of automatic speech recognition to convert speech-signal into text comprising one or more data processing apparatus;

and a computer-readable medium coupled to the one or more data processing apparatus having instructions stored thereon which, when executed by the one or more data processing apparatus, cause the one or more data processing apparatus to perform a method comprising:
A) segmenting the speech signal into pitch-synchronous frames, wherein for voiced sections each said frame is a single pitch period;
B) for each frame, equalizing the two ends of the waveform using an ends-matching program;
C) generating an amplitude spectrum of each said frame using Fourier analysis;
D) transforming each said amplitude spectrum into a timbre vector using Laguerre functions;
E) performing acoustic decoding to find a list of most likely phonemes or sub-phoneme units for each said timbre vector by comparing with a timbre vector database;
F) decoding the sequence of the list of the most likely phonemes or sub-phoneme units using a language-model database to find out the most likely text;
wherein the segmenting of the speech-signal is based on an analysis of the speech signals using an asymmetric window including:
a) conducting, for a speaker, a test to find the best size of the asymmetric window;
b) convoluting the speech-signal with the said asymmetric window to form a profile function;
c) picking up the maxima in the said profile function as segmentation points;
d) extend the segmentation points to unvoiced sections.

9 more elements

Dependencies are inferred from the claim text.

Facts

Application: 14/216,684Filed 17 Mar 2014 · continuation in part
Priority: 3 Dec 2012
Published: US 2014/0088958 A117 Jul 2014
Granted: 27 Jan 2015
Term ends: Not estimatedThe source records don’t support an estimate.

Classification4 codes

G10L 15/02
Speech recognition: feature extraction for speech recognition; Selection of recognition unit
G10L 25/90
Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00: pitch determination of speech signals
G10L 25/93
Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00: discriminating between voiced and unvoiced…
G10L
Speech analysis techniques or speech synthesis

Family

Application13/692,584Parent of this continuation-in-part · filed 3 Dec 2012
US 2014/0200889 A1Pre-grant publication · published 17 Jul 2014
US 8,719,030 B2Parent of this continuation-in-part
US 2014/0088958 A1Published application
US 8,942,977 B2This patent

From the US filing history. Foreign family members aren’t listed.