Speech processor

• US 4,912,766 A
• Filed: 06/01/1987
• Issued: 03/27/1990
• Est. Priority Date: 06/02/1986
• Status: Expired due to Term

• Current Assignee: British Telecommunications PLC (BT Group PLC)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Forse, Nicholas J. A.
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US07/055,966
• Time in Patent Office: 1,030 Days
• Field of Search: 381/29-48, 364/513.5, 370/70, 375/98
• US Class Current: 704/225
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• 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 2021/02168 the estimation exclusively taking place during speech pauses
• CPC Class Codes: G10L 25/18 the extracted parameters being spectral information of each sub-band
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10

Abstract

In a speech processor such as a speech recognizer, the problem of distortion of extracted features caused by adaptation of the input automatic gain control (AGC) during feature extraction is solved by storing the AGC'"'"'s gain coefficient along with the energy level of each extracted feature. At the end of the sampling period the stored gain coefficients are set equal to the minimum stored coefficient and the associated energy levels adjusted accordingly. The AGC circuit may comprise a digitally switched attenuator under the control of a microprocessor performing the speech recognition.

Claims

1. A speech processor comprising:

an input to receive speech signals;

signal processing means connected to extract spectral parameters from said speech signals;

an analogue to digital converter connected to digitise said extracted parameters over a sampling period;

an automatic gain control means connected to control the signal level applied to said converter;

means to store, at least temporarily, the spectral parameters and, for each such stored parameter, to also store a gain coefficient indicative of the gain applied by the gain control means; and

adjustment means operable at intervals to perform an adjustment process in which the gain coefficients stored in said period are, if different, set equal to the lowest gain coefficient stored since the previous adjustment period, the magnitudes of the corresponding stored spectral parameters also being adjusted proportionally.

2. A speech processor as in claim 1 wherein said adjustment means includes control means connected to cause each extracted spectral parameter to correspond to the energy content of a particular frequency band in a time slot of length t, and to determine for each extracted parameter, the signal level applied to the analogue to digital converter in a small fraction of the time t, and if the signal level is greater than a predetermined level reducing the gain and re-assessing the signal level, the signal level assessment and the gain reduction being repeated within time slot t until the signal level is at a finalised level not exceeding said predetermined level.

3. A speech processor as in claim 2 wherein said control means causes said predetermined level to be equal to the maximum level which does not exceed the dynamic range of the analogue to digital converter.

4. A speech processor as in claim 2 or claim 3 wherein said control means causes spectral parameters to be established in a single time slot of length t for a plurality of discrete frequency bands, and causes the different frequency bands to be addressed sequentially, with the finalised gain coefficient of any frequency band being used as the initial gain coefficient of the next addressed frequency band.

5. A speech processor as in claim 2 or 3 wherein the control means causes the sampling period to be made up of a plurality of time slots of length t.

6. A speech processor as in claim 1, 2, or 3, configured as a speech recogniser.

7. A speech processor as in claim 1, 2 or 3, wherein the control means comprises a microprocessor and the gain control means comprises a digitally switched attenuator under the control of the microprocessor having an input connected to a digitised output of the analogue to digital converter, the gain of the attenuator being determined by the microprocessor.

8. A cellular radio terminal comprising:

a speech recogniser for selecting functions in response to voiced instructions, characterised in that the speech recogniser comprises a speech processor as claimed in claim 1, 2 or 3.

9. A speech processor comprising:

an input for receiving speech signals;

signal processing means for extracting spectral parameters from said speech signals;

automatic gain control means for controlling the magnitudes of said parameters;

means for storing, at least temporarily, said gain-controlled parameters and gain coefficients indicative of the gains applied to the parameters by said gain control means; and

adjustment means operative at intervals to perform an adjustment process comprising examining the coefficients relating to the parameters stored since the previous such adjustment process and, if they are different, adjusting the parameter magnitudes such that all those parameters have magnitudes corresponding to a single value of gain coefficient.

10. A speech processor comprising:

filtering means connected to receive an input speech signal for separating said speech signal into a plurality of spectral components;

amplifier means for sequentially amplifying said plurality of spectral components by corresponding signal gain factors;

converter means connected to receive said amplified speech signal for converting said amplified plurality of spectral components into digital signals representing said spectral components;

means for measuring said signal gain factors;

means connected to said measuring means and to said converter means for storing said digital signals representing said spectral components and for also storing said measured gain factors corresponding to said spectral components; and

means for adjusting the value of stored spectral components associated with at least an initial portion of said speech signal so as to be based on different signal gain factors derived, at least in part, on gain factors used to derive subsequent spectral components of said speech signal.

11. A processor as in claim 10 wherein said means for adjusting produces outputs spectral component data for the entire input speech signal referenced to a single common gain factor.

12. A speech processor comprising:

spectral analysis means connected to receive input speech signals and to provide measured spectral parameter data therefrom and including an automatic gain adjustment means for adjusting the level of speech signal components being measured but which, upon the initial portion of a speech signal may inherently provide excessive signal gain which gain is thereafter automatically adjusted to lesser levels during the continuing speech signal occurrence; and

data storage and adjusting means for storing and subsequently reducing excessive signal gain factors associated with the speech signal components during at least said initial portion of a speech signal to produce an approximately uniform effective signal gain over the entire speech signal occurrence thereby producing more reliably recognizable measured spectral parameter data.

13. A speech processor comprising:

a speech signal input;

plural frequency band pass filter channels connected to receive speech signals from said input and to respectively provide spectral outputs representing the input speech signal content within corresponding frequency bands;

a multiplexer connected to receive said spectral outputs and to select one such spectral output as a function of first digital control signals supplied thereto;

a digitally controlled automatic gain control circuit connected to receive said selected spectral output and to apply a selected gain factor thereto as a function of a second digital control signals supplied thereto;

an analog-to-digital converter connected to receive gain-adjusted signals from the gain control circuit and to supply digitized spectral signals corresponding thereto; and

a digital control unit connected to supply said first and second digital control signals so as to periodically sample and gain adjust said spectral outputs and including a digital data memory connected to receive and store both (a) said digitized spectral signals and (b) the gain factor used in the automatic gain control circuit to generate such digitized spectral signals;

said digital control unit including means for subsequently adjusting the values of digital spectral signals corresponding to an initial portion of a speech signal so as to be based on a lower gain factor derived, at least in part, from gain factors utilized for later occurring portions of the speech signal.

14. A speech processing method comprising:

filtering an input speech signal into a plurality of spectral components;

sequentially amplifying said plurality of spectral components by corresponding signal gain factors;

converting said amplified plurality of spectral components into digital signals representing said spectral components;

measuring and storing said digital signals representing (a) spectral components of an input speech signal and (b) said corresponding signal gain factors; and

adjusting the value of stored spectral components associated with at least an initial portion of said speech signal so as to be based on different signal gain factors used to derive subsequent spectral components of said speech signal.

15. A speech processing method as in claim 14 wherein said adjusting step includes producing output spectral component data for the entire input speech signal referenced to a single common gain factor.

16. A speech processing method comprising:

receiving input speech signals and providing measured spectral parameter data therefrom including an automatic gain adjustment adjusting the level of speech signal components being measured but which, upon the initial portion of a speech signal may inherently provide excessive signal gain which gain is thereafter automatically adjusted to lesser levels during the continuing speech signal occurrence; and

storing and subsequently reducing excessive signal gain factors associated with the speech signal components during at least said initial portion of a speech signal to produce an approximately uniform effective signal gain over the entire speech signal occurrence thereby producing more reliably recognizable measured spectral parameter data.

17. A speech processing method comprising:

receiving input speech signals and providing plural spectral outputs representing the input speech signal content within corresponding frequency bands;

receiving said spectral outputs and selecting one such spectral output as a function of first digital control signals;

applying a selected gain factor to said selected spectral output as a function of second digital control signals;

converting said gain-adjusted signals from the gain control circuit to digitized spectral signals corresponding thereto;

supplying said first and second digital control signals to successively sample and gain adjust said spectral outputs and storing both (a) said digitized spectral signals and (b) the gain factor used in the automatic gain control circuit to generate such digitized spectral signals; and

subsequently adjusting the values of digital spectral signals corresponding to an initial portion of a speech signal so as to be based on a lower gain factor derived, at least in part, from gain factors utilized for later occurring portions of the speech signal.

18. A speech processing method comprising:

(a) storing successive spectral parameters derived from periodically sampling an input speech signal which has been subjected to automatic gain control;

(b) also storing gain data representing the gain factors used to successively derive such stored spectral parameters; and

(c) subsequently adjusting the stored spectral parameters representing an initial portion of an input speech signal to reduced values in compensation for an initially excessive gain factor.

19. A speech processing method as in claim 18 wherein said adjusting step changes said stored spectral parameters to values which would have been recorded if a uniform constant gain factor had been used for the entire speech signal.