Method for reduced computation of T-matrix training for speaker recognition
A system and method for improving T-matrix training for speaker recognition, comprising receiving an audio input, divisible into a plurality of audio frames including at least an audio sample of a human speaker; generating for each audio frame a feature vector; generating for a first plurality of feature vectors centered statistics of at least a zero order and a first order; generating a first i-vector, the first i-vector representing the human speaker; and generating an optimized T-matrix training sequence computation, based on at least the first i-vector.
1. A method for improving T-matrix training for speaker recognition, comprising:
receiving an audio input, divisible into a plurality of audio frames, including at least an audio sample of a human speaker;
generating for each audio frame a feature vector;
generating for a first plurality of feature vectors centered statistics of at least a zero order and a first order;
generating a first i-vector, the first i-vector representing the human speaker; and
generating an optimized T-matrix training sequence computation, based on: at least the first i-vector, an initialized T-matrix, the centered statistics, and a Gaussian Mixture Model (GMM) of a trained universal background model (UBM), wherein generating the optimized T-matrix training sequence further comprises:
generating a first multi-dimensional array comprising a plurality of duplicated matrices, wherein each matrix includes a plurality of GMM mean vectors;
generating a multi-dimensional feature matrix, comprising a plurality of feature matrices, wherein each feature matrix corresponding to a feature vector of a single audio frame; and
generating a multi-dimensional delta array based on the first multi-dimensional array and the multi-dimensional feature matrix.
2. The method of claim 1 , wherein the audio sample has a length above a threshold.
3. The method of claim 1 , wherein the GMM is associated with a plurality of components, each defined by a covariance matrix, a mean vector, and a weight.
4. The method of claim 1 , wherein generating the optimized T-matrix training sequence further comprises:
generating a feature matrix based on the feature vector;
generating a GMM mean matrix based on a plurality of mean GMM vectors associated with the plurality of GMM components; and
generating a delta matrix based on the feature matrix and the GMM mean matrix.
5. The method of claim 1 , wherein generating the optimized T-matrix training sequence further comprises:
detecting diagonal matrices, and performing only those computations that involve diagonal elements.
6. The method of claim 1 , wherein generating the optimized T-matrix training sequence further comprises:
detecting computations in an intermediate result that generate an off diagonal element of a matrix which is diagonalized; and
eliminating the computation of the intermediate result.
7. The method of claim 1 , wherein generating the optimized T-matrix training sequence further comprises:
detecting a recurring computation; and
precomputing the recurring computation and storing a precomputed result in a cache of a processor.
8. The method of claim 1 , wherein generating the optimized T-matrix training sequence further comprises:
detecting a computation between a first matrix and a second matrix; and
replacing the detected computation with an element-by-element computation between the first matrix and the second matrix, in response to determining that a replacement will result in a same output.
9. The method of claim 1 , wherein a plurality of i-vectors is generated, wherein each i-vector corresponds to a single human speaker.
10. The method of claim 9 , wherein a first group of the plurality of i-vectors are generated from a first audio input, and a second group is generated from a second audio input.
11. The method of claim 1 , further comprising:
repeating the method for a plurality of iterations based on a single audio frame.
12. The method of claim 1 , further comprising:
repeating the method for each of the plurality of audio frames.
13. The method of claim 1 , wherein the method is performed in multiple iterations on each audio frame.
14. The method of claim 1 , wherein generating the optimized T-matrix training sequence further comprises:
generating score between a first i vector and a second i vector; and
repeating for plurality of i vectors.
15. The method of claim 1 , wherein the feature vector is an MFCC vector.
16. The method of claim 1 , further comprising:
executing the optimized T-matrix training sequence on at least one of: a general-purpose graphics processor unit (GPGPU), and a multi-core CPU.
17. A non-transitory computer readable medium having stored thereon instructions for causing a processing circuitry to execute a process, the process comprising:
receiving an audio input, divisible into a plurality of audio frames, including at least an audio sample of a human speaker;
generating for each audio frame a feature vector;
generating for a first plurality of feature vectors centered statistics of at least a zero order and a first order;
generating a first i-vector, the first i-vector representing the human speaker; and
generating an optimized T-matrix training sequence computation, based on: at least the first i-vector, an initialized T-matrix, the centered statistics, and a Gaussian Mixture Model (GMM) of a trained universal background model (UBM), wherein generating the optimized T-matrix training sequence further comprises:
generating a first multi-dimensional array comprising a plurality of duplicated matrices, wherein each matrix includes a plurality of GMM mean vectors;
generating a multi-dimensional feature matrix, comprising a plurality of feature matrices, wherein each feature matrix corresponding to a feature vector of a single audio frame; and
generating a multi-dimensional delta array based on the first multi-dimensional array and the multi-dimensional feature matrix.
18. A system for improving T-matrix training for speaker recognition, comprising:
a processing circuitry; and
a memory, the memory containing instructions that, when executed by the processing circuitry, configure the system to:
receive an audio input, divisible into a plurality of audio frames including at least an audio sample of a human speaker;
generate for each audio frame a feature vector;
generate for a first plurality of feature vectors centered statistics of at least a zero order and a first order;
generate a first i-vector, the first i-vector representing the human speaker; and
generate an optimized T-matrix training sequence computation, based on: at least the first i-vector, an initialized T-matrix, the centered statistics, and a Gaussian Mixture Model (GMM) of a trained universal background model (UBM), wherein generating the optimized T-matrix training sequence further comprises:
generating a first multi-dimensional array comprising a plurality of duplicated matrices, wherein each matrix includes a plurality of GMM mean vectors;
generating a multi-dimensional feature matrix, comprising a plurality of feature matrices, wherein each feature matrix corresponding to a feature vector of a single audio frame; and
generating a multi-dimensional delta array based on the first multi-dimensional array and the multi-dimensional feature matrix.