Quantization using frequency and mean compensated frequency input data for robust speech recognition

• US 6,418,412 B1
• Filed: 08/28/2000
• Issued: 07/09/2002
• Est. Priority Date: 10/05/1998
• Status: Expired due to Term

• Current Assignee: RPX Corporation
• Original Assignee: Legerity Incorporated (Microsemi Semiconductor Corp.)
• Inventors: Cong, Lin, Asghar, Safdar M.
• Primary Examiner(s): {haeck over (S)}mits, Ta̅livaldis Ivars
• Assistant Examiner(s): ARMSTRONG, ANGELA A
• Application Number:
  US09/649,737
• Time in Patent Office: 680 Days
• Field of Search: 704/205, 704/219, 704/222, 704/226, 704/230, 704/231, 704/232, 704/233, 704/240, 704/256, 704/243
• US Class Current: 704/256.5
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/144 Training of HMMs
• 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

Abstract

A speech recognition system utilizes multiple quantizers to process frequency parameters and mean compensated frequency parameters derived from an input signal. The quantizers may be matrix and vector quantizer pairs, and such quantizer pairs may also function as front ends to a second stage speech classifiers such as hidden Markov models (HMMs) and/or utilizes neural network postprocessing to, for example, improve speech recognition performance. Mean compensating the frequency parameters can remove noise frequency components that remain approximately constant during the duration of the input signal. HMM initial state and state transition probabilities derived from common quantizer types and the same input signal may be consolidated to improve recognition system performance and efficiency. Matrix quantization exploits the “evolution” of the speech short-term spectral envelopes as well as frequency domain information, and vector quantization (VQ) primarily operates on frequency domain information. Time domain information may be substantially limited which may introduce error into the matrix quantization, and the VQ may provide error compensation. The matrix and vector quantizers may split spectral subbands to target selected frequencies for enhanced processing and may use fuzzy associations to develop fuzzy observation sequence data. A mixer may provide a variety of input data to the neural network for classification determination. Fuzzy operators may be utilized to reduce quantization error. Multiple codebooks may also be combined to form single respective codebooks for split matrix and split vector quantization to reduce processing resources demand.

Claims

1. A signal recognition system comprising:

a frequency parameter mean compensation module to receive frequency parameters of an input signal and to generate mean compensated frequency parameters from the received input signal frequency parameters;

a first quantizer to receive the input signal frequency parameters and to quantize the input signal frequency parameters;

a second quantizer to receive the input signal mean compensated frequency parameters and to quantize the input signal mean compensated frequency parameters; and

a backend processor to receive the quantized input signal frequency parameters and the input signal mean compensated input signal frequency parameters and to generate an input signal classification therefrom.

2. The signal recognition system as in claim 1 wherein for TO samples of the input signal, an i^th frequency parameter of the input signal in a j^th sample of the input signal, j=1, 2, . . . , TO, is represented by s(i)_j, and each mean compensated frequency parameter, s(i)_j,(m) is generated by the frequency parameter mean compensation module in accordance with:

${s<br><br>\left(i\right)}_{j,m<br><br>\text{\hspace{1em}<br><br>}<br><br>c}={s<br><br>\left(i\right)}_j-\frac{1}{\mathrm{TO}}<br><br>\underset{j=1}{\overset{\mathrm{TO}}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{s<br><br>\left(i\right)}_j.$

3. The signal recognition system as in claim 1 wherein the first quantizer comprises a vector quantizer and a matrix quantizer.

4. The signal recognition system as in claim 3 wherein the vector quantizer is capable of generating vector quantized data and the matrix quantizer is capable of generating matrix quantized data, wherein the vector quantized data is capable of being combined with the matrix quantized data to generate the quantized input signal frequency parameters.

5. The signal recognition system as in claim 1 wherein the second quantizer comprises a vector quantizer and a matrix quantizer.

6. The signal recognition system as in claim 5 wherein the vector quantizer is capable of generating mean compensated vector quantized data and the matrix quantizer is capable of generating mean compensated matrix quantized data, wherein the mean compensated vector quantized data is capable of being combined with the mean compensated matrix quantized data to generate the quantized input signal mean compensated frequency parameters.

7. The signal recognition system as in claim 1 wherein the backend processor comprises a first group of hidden Markov models to receive quantized training output data from the first quantizer and a second group of hidden Markov models to receive quantized training output data from the second quantizer.

8. The signal recognition system as in claim 7 further comprising:

a first stochastic module to receive quantized training output data from the first quantizer to determine the respective probabilities of each of the first group of hidden Markov models; and

a second stochastic module to receive quantized training output data from the second quantizer to determine the respective probabilities of each of the second group of hidden Markov models.

9. The signal recognition system as in claim 8 wherein the stochastic module comprises a Viterbi algorithm.

10. The signal recognition system as in claim 1 wherein the backend processor comprises a neural network to receive respective quantized output data from the first and second quantizers.

11. The signal recognition system as in claim 1 further comprising:

a memory having code to implement the frequency parameter compensation module, the first quantizer, the second quantizer, and the backend processor; and

a processor coupled to the memory to execute the code.

12. A method comprising the steps of:

sampling an input signal having a noise component;

characterizing the sampled input signal with frequency parameters;

generating mean compensated frequency parameters from the frequency parameters to substantially remove the noise component;

providing the frequency parameters to a first quantizer;

providing the mean compensated frequency parameters to a second quantizer; and

quantizing the frequency parameters with the first quantizer to generate first quantization data;

quantizing the mean compensated frequency parameters with the second quantizer to generate second quantization data; and

providing the first and second quantization data to a backend processor to classify the input signal.

13. The method as in claim 12 wherein an i^th frequency parameter of the sampled input signal in a j^th sample of the input signal, j=1, 2, . . . , TO, is represented by s(i)_j and the step of generating mean compensated frequency parameters from the frequency parameters comprises the step of:

mean compensating each of D frequency parameters for TO frames of the sampled input signal wherein the i^th mean compensated frequency parameter, s(i)_j,(m), i=1, 2, . . . , D, is generated in accordance with;

${s<br><br>\left(i\right)}_{j,m<br><br>\text{\hspace{1em}<br><br>}<br><br>c}={s<br><br>\left(i\right)}_j-\frac{1}{\mathrm{TO}}<br><br>\underset{j=1}{\overset{\mathrm{TO}}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{s<br><br>\left(i\right)}_j.$

14. The method as in claim 12 wherein the step of quantizing the frequency parameters with the first quantizer comprises the steps of:

quantizing the frequency parameters with a matrix quantizer;

quantizing the frequency parameters with a vector quantizer; and

combining the frequency parameters quantized with the matrix quantizer with the frequency parameters quantized with the vector quantized data to generate the first quantization data.

15. The method as in claim 12 wherein the step of quantizing the mean compensated frequency parameters with the second quantizer comprises the steps of:

quantizing the mean compensated frequency parameters with a matrix quantizer;

quantizing the mean compensated frequency parameters with a vector quantizer; and

combining the mean compensated frequency parameters quantized with the matrix quantizer with the mean compensated frequency parameters quantized with the vector quantized data to generate the second quantization data.

16. The method as in claim 12 wherein the step of providing the first and second quantization data to a backend processor comprises the steps of:

providing the first and second quantization data to a stochastic module having access to data from a plurality of hidden Markov models; and

utilizing the stochastic module to determine classification probabilities from each of the respective hidden Markov models.

17. The method as in claim 16 wherein the step of providing the first and second quantization data to a backend processor further comprises the step of:

providing the classification probabilities from each of the respective hidden Markov models to a neural network.

18. A signal recognition system comprising:

a frequency parameter mean compensation module to receive frequency parameters of an input signal having a noise component and to generate mean compensated frequency parameters from the received input signal frequency parameters in order to substantially remove the noise component;

a first quantizer to receive the input signal frequency parameters and to quantize the input signal frequency parameters;

a second quantizer to receive the input signal mean compensated frequency parameters and to quantize the input signal mean compensated frequency parameters; and

a backend processor to receive the quantized input signal frequency parameters and the input signal mean compensated input signal frequency parameters and to generate an input signal classification therefrom.

19. The signal recognition system as in claim 18, wherein for TO samples of the input signal, an i^th frequency parameter of the input signal in a j^th sample of the input signal, j=1, 2, . . . , TO, is represented by s(i)_j, and each mean compensated frequency parameter, s(i)_j,(m) is generated by the frequency parameter mean compensation module in accordance with:

${s<br><br>\left(i\right)}_{j,m<br><br>\text{\hspace{1em}<br><br>}<br><br>c}={s<br><br>\left(i\right)}_j-\frac{1}{\mathrm{TO}}<br><br>\underset{j=1}{\overset{\mathrm{TO}}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{s<br><br>\left(i\right)}_j.$

20. The signal recognition system as in claim 18 wherein the first quantizer comprises a vector quantizer and a matrix quantizer.

21. The signal recognition system as in claim 20 wherein the vector quantizer is capable of generating vector quantized data and the matrix quantizer is capable of generating matrix quantized data, wherein the vector quantized data is capable of being combined with the matrix quantized data to generate the quantized input signal frequency parameters.

22. The signal recognition system as in claim 18 wherein the second quantizer comprises a vector quantizer and a matrix quantizer.

23. The signal recognition system as in claim 22 wherein the vector quantizer is capable of generating mean compensated vector quantized data and the matrix quantizer is capable of generating mean compensated matrix quantized data, wherein the mean compensated vector quantized data is capable of being combined with the mean compensated matrix quantized data to generate the quantized input signal mean compensated frequency parameters.

24. The signal recognition system as in claim 18 wherein the backend processor comprises a first group of hidden Markov models to receive quantized training output data from the first quantizer and a second group of hidden Markov models to receive quantized training output data from the second quantizer.

25. The signal recognition system as in claim 24 further comprising:

a first stochastic module to receive quantized training output data from the first quantizer to determine the respective probabilities of each of the first group of hidden Markov models; and

a second stochastic module to receive quantized training output data from the second quantizer to determine the respective probabilities of each of the second group of hidden Markov models.

26. The signal recognition system as in claim 25 wherein the stochastic module comprises a Viterbi algorithm.

27. The signal recognition system as in claim 18 wherein the backend processor comprises a neural network to receive respective quantized output data from the first and second quantizers.

28. The signal recognition system as in claim 18 further comprising:

a memory having code to implement the frequency parameter compensation module, the first quantizer, the second quantizer, and the backend processor; and

a processor coupled to the memory to execute the code.