METHOD OF CORRECTING ERRORS IN BINARY MASKS

• US 20110051948A1
• Filed: 08/04/2010
• Published: 03/03/2011
• Est. Priority Date: 08/26/2009
• Status: Active Grant

• Current Assignee: Oticon A/S (William Demant Fonden)
• Original Assignee: Oticon A/S (William Demant Fonden)
• Inventors: Kjems, Ulrik, Pedersen, Michael Syskind, Christensen, Mads Graesbøll, BOLDT, Jesper Bünsow, Jensen, Søren Holdt
• Granted Patent: US 8,626,495 B2
• Time in Patent Office: Days
• US Class Current: 381/73.1
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2021/065 Aids for the handicapped in understanding
• CPC Class Codes: G10L 21/02 Speech enhancement, e.g. noise reduction or echo cancellation reducing echo effects in line transmission systems H04B3/20; echo suppression in hands-free telephones H04M9/08
• CPC Class Codes: G10L 25/18 the extracted parameters being spectral information of each sub-band
• CPC Class Codes: G10L 25/48 specially adapted for particular use
• CPC Class Codes: H04R 2225/43 Signal processing in hearing aids to enhance the speech intelligibility
• CPC Class Codes: H04R 25/505 using digital signal processing

Abstract

The invention relates to a method of identifying and correcting errors in a noisy binary mask. An object of the present invention is to provide a scheme for improving a binary mask representing speech. The problem is solved in that the method comprises a) providing a noisy binary mask comprising a binary representation of the power density of an acoustic signal comprising a target signal and a noise signal at a predefined number of discrete frequencies and a number of discrete time instances; b) providing a statistical model of a clean binary mask representing the power density of the target signal; and c) using the statistical model to detect and correct errors in the noisy binary mask. This has the advantage of providing an alternative and relatively simple way of improving an estimate of a binary mask representing a speech signal. The invention may e.g. be used for speech processing, e.g. in a hearing instrument.

Claims

1. A method of identifying and correcting errors in a noisy binary mask, the method comprising

a) providing a noisy binary mask comprising a binary representation of the power density of an acoustic signal comprising a target signal mixed with a noise signal at a predefined number of discrete frequencies and a number of discrete time instances;

b) providing a statistical model of a clean binary mask representing the power density of the target signal; and

c) using the statistical model to detect and correct errors in the noisy binary mask.

2. A method according to claim 1 wherein the statistical model is based on Hidden Markov Models.

3. A method according to claim 1 wherein the step of providing a statistical model comprises providing a training set of a clean binary mask comprising a binary representation of power density of the target signal at the predefined number of discrete frequencies and a number of discrete time instances

4. A method according to claim 1 wherein the step of providing a statistical model comprises providing a training set of a noisy binary mask comprising a binary representation of the power density of a mixture of the target signal and a noise signal at the predefined number of discrete frequencies and a number of discrete time instances.

5. A method according to claim 1 wherein the statistical model comprises states and observations and wherein the states are constituted by vectors representing the target signal at the predefined number of discrete frequencies at a number of points in time, and wherein the observations are constituted by vectors representing the target signal mixed with the noise signal at the predefined number of discrete frequencies at a number of points in time.

6. A method according to claim 5 wherein each state is constituted by a binary vector Q_t representative of the target signal at the predefined number of discrete frequencies in a given time unit t.

7. A method according to claim 5 wherein for each state Q_t a corresponding observation X_t is constituted by a vector comprising the probability of a one for each of the predefined number of discrete frequencies, given the state in question.

8. A method according to claim 5 wherein state transition probabilities defining a probability of changing from state Q_t to Q_t+1 are provided, and e.g. arranged in a square matrix A with size N×

N, where N is the total number of states and where the matrix element at the m^th row and n^th column in the matrix represents the probability of changing from state m to state n.

9. A method according to claim 7 wherein the observation probabilities are arranged in a matrix B with size F×

N, where F is the number of frequencies and N is the total number of states and where the matrix element at the p^th row and q^th column in the matrix represents the probability of a one at the p^th frequency of the q^th state.

10. A method according to claim 5 comprising a quantization of the clean binary mask wherein a subset of N_q states are determined, where N_q<

N, each state being constituted by a binary mode vector representative of the target signal at the predefined number of discrete frequencies.

11. A method according to claim 10 wherein the state transition probabilities a_ij of the matrix A are calculated from the quantized binary mask by counting the number n_ij of state changes Qi to Qj, where i=1, 2, . . . , N_q and j=1, 2, . . . , N_q and divide each number n_ij by the total number N_sc,i of state changes of the quantized binary mask from state i.

12. A method according to claim 10 wherein the observation probabilities B(k,n) of the matrix B are calculated based on two contributions B_c and b_m according to the formula B(k,n)=B_c(k,n)+b_n(k)−

B_c(k,n)·

b_n(k), where f=1, 2, . . . , F is the frequency index and k=1, 2, . . . , N_q is the state index, where B_c(k,n) are the observation probabilities defined by the clean binary mask and the quantized clean binary mask, and where the observation probabilities b_n(k) defining the probability of ones at each frequency generated by the noise signal.

13. A method according to claim 12 where the observation probabilities b_n(k) are calculated from characteristics of the of the noise signal in question.

14. A method according to claim 13 where the observation probabilities b_n(k) are calculated during time periods where no voice signals are present.

15. A method according to claim 12 where the observation probabilities are calculated based on a recordal of the noise signal alone prior to the use of the statistical model to detect and correct errors in the noisy binary mask.

16. A method according to claim 1 where a target binary mask TBM_est is estimated from the following formula:

17. A method according to claim 1 wherein an estimate of a noise-free binary mask is determined from the noisy binary mask as the most probable sequence of states using the Viterbi algorithm, the states being constituted by vectors representing the target signal in the binary domain at the predefined number of discrete frequencies at a number of points in time.

18. A method according to claim 17 wherein the most probable sequence of states are calculated based on a number of past, the present, and a number of future observations, the observations being constituted by binary vectors representing the target signal mixed with the noise signal at the predefined number of discrete frequencies at a number of points in time.

19. A method according to claim 1 wherein the noisy binary mask represents speech in noise.

20. A tangible computer-readable medium storing a computer program comprising program code means for causing a data processing system to perform the steps of the method of claim 1, when said computer program is executed on the data processing system.

21. A data processing system comprising a processor and program code means for causing the processor to perform the steps of the method of claim 1.