Method and apparatus for recognizing deformed speech

• US 6,006,180 A
• Filed: 01/27/1995
• Issued: 12/21/1999
• Est. Priority Date: 01/28/1994
• Status: Expired due to Term

• Current Assignee: Orange S.A.
• Original Assignee: Orange S.A.
• Inventors: Chollet, Gerard, Bardaud, Philippe
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/379,870
• Time in Patent Office: 1,789 Days
• Field of Search: 395/2.18, 395/2.77, 395/2.28, 395/2.74, 395/2.32, 395/2.73, 395/2.67, 395/2.81, 704/209, 704/268, 704/219, 704/265, 704/223, 704/231, 704/258, 704/236, 704/243, 704/261, 704/264, 704/266
• US Class Current: 704/223
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• CPC Class Codes: G10L 2021/03643 Diver speech

Abstract

An apparatus and method for recognizing deformed speech signals outputted from a microphone includes a module for comparing deformed signals with simulated deformed signals that are generated from non-deformed speech signals that have previously been digitized and stored in a memory. Frequencies of the formants of simulated deformed signals are about two-three times the frequencies of the formants of digitized non-deformed signals.

Claims

1. A system for processing a non-deformed speech signal spectrum, the speech signal spectrum having a first and a second frequency which represent a respective formant of said speech signal, the system comprising:

extraction means responsive to said speech signal to extract therefrom an excitation signal representative of the sound and vibration sources of the speech;

envelope determination means responsive to said speech signal to compute coefficients characteristic of the shape of the spectrum envelope of said speech signal;

interpolation means responsive to said excitation signal to generate an interpolated excitation signal having a waveform that is identical to the waveform of said excitation signal and having a point density that is about two to three times the point density of said excitation signal;

synthesizer means responsive to said interpolated excitation signal and said characteristic coefficients to synthesize a deformed speech signal;

electronic means for linearly increasing the frequencies of the formants of said speech signal by a factor of about 2 to 3.

2. The system according to claim 1, further including a linear predicting coding module which combines said extraction means and said envelope determining means.

3. The system according to claim 1, further including preprocessor means for preprocessing said speech signal comprising

pre-emphasis means for boosting the high frequency components of said speech signal and

windowing means for weighting a signal segment in application of a curve of a predetermined shape.

4. A system for recognizing deformed speech signals delivered by a microphone, the system comprising:

an apparatus for generating speech data representative of simulated deformed signals generated from non-deformed analog speech signals and for storing said data in a memory, said apparatus comprising;

a) conversion means for digitizing said analog speech signals into a time sequence of digital values representing a digitized sampled speech signal, said digitized sampled speech signal having high frequency component that represent said formants of speech;

b) digital pre-emphasis means for boosting the high frequency components of the digitized sampled speech signal;

c) windowing means for weighting a window in application of a curve of predetermined shape;

d) extraction means responsive to said speech data representative of said speech signal to extract digital excitation data representative of an excitation signal;

e) envelope determination means responsive to said speech data to compute coefficients characteristic of a shape of a spectrum envelope of said digitized speech signal;

f) interpolation means responsive to said excitation data to generate interpolated excitation data having a waveform that is identical to a waveform of said excitation data and having a point density that is about two to three times a point density of said excitation data; and

g) synthesizer means responsive to said interpolated excitation data and to said characteristic coefficients to synthesize data representative of a simulated deformed speech signal;

means for converting said simulated deformed speech data into analog simulated deformed speech signals, wherein the frequencies of the formants of said simulated deformed signal are about two to three times the frequencies of the formants of said digitized non-deformed signals; and

a comparator module for comparing the deformed speech signals with said simulated deformed signals.

5. A method of recognizing deformed speech comprising the steps of:

digitizing and storing non-deformed speech signals in a memory;

generating simulated deformed speech signals from said non-deformed speech signals, the simulated deformed speech signals including frequencies representing formants of speech;

increasing the frequencies of the formants of said simulated deformed signals by a factor of about two to three times with respect to the frequencies of the formants of said digitized non-deformed signals; and

comparing said deformed signals with said simulated deformed signals.

6. The method according to claim 5, wherein

said signal is sampled and digitized at a first frequency, the resulting successive data values being stored in a memory,

said signal obtained by subjecting data representative of said signal to digital-to-analog conversion and sampling at a second frequency by a factor of two to three times said first frequency,

said data representative of said signal obtained by one of synthesis and superposition of data representative of an interpolation excitation signal and a spectrum envelope defined by coefficients

wherein said data representative of said interpolation excitation signal is obtained by interpolation of data representative of a non-interpolated excitation signal and wherein

said data representative of said non-interpolated excitation signal and said characteristic coefficients of the spectrum envelope are calculated from said non-deformed speech signal by a method of linear predictive coding.

7. A method according to claim 5, wherein said simulated deformed signals are generated by using a multiple linear regression method applied to the cepstre vectors of said non-deformed speech signals.

8. The method according to claim 5, wherein comparing said deformed signals with said simulated deformed signals is performed by computing distances on said curve.