System and method for speaker change detection
A method for training a neural network of a neural network based speaker classifier for use in speaker change detection. The method comprises: a) preprocessing input speech data; b) extracting a plurality of feature frames from the preprocessed input speech data; c) normalizing the extracted feature frames of each speaker within the preprocessed input speech data with each speaker's mean and variance; d) concatenating the normalized feature frames to form overlapped longer frames having a frame length and a hop size; e) inputting the overlapped longer frames to the neural network based speaker classifier; and f) training the neural network through forward-backward propagation.
1. A method for training a neural network of a neural network based speaker classifier for use in speaker change detection, the method comprising:
a) preprocessing input speech data;
b) extracting a plurality of feature frames from the preprocessed input speech data;
c) normalizing the extracted feature frames of each speaker within the preprocessed input speech data with each speaker's mean and variance;
d) concatenating the normalized feature frames to form overlapped longer frames having a frame length and a hop size;
e) inputting the overlapped longer frames to the neural network based speaker classifier; and
f) training the neural network through forward-backward propagation, wherein step (a) comprises:
a.1) scaling a maximum of absolute amplitude of the input speech data to one;
a.2) performing voice activity detection on the scaled input speech data; and
a.3) eliminating unvoiced portions of the scaled input speech data, and
wherein step (a.2) comprises:
a.2.1) segmenting the scaled input speech data into overlapped frames with a window size and a hop size;
a.2.2) determine a short-term energy E of each frame;
a.2.3) determine a spectral centroid C of each frame;
a.2.4) remove any frame from the scaled input speech data in which the short-term energy E of the frame is below a predetermined threshold TE; and
a.2.5) remove any frame from the scaled input speech data in which the spectral centroid C of the frame is below a predetermined threshold TC.
2. The method of claim 1 , wherein the window size of the overlapped frames is 50 ms and the hop size of the overlapped frames is 25 ms.
3. The method of claim 1 , wherein the short-term energy E comprises:
E
=
1
N
∑
n
=
1
N
s
(
n
)
2
,
where s(n) is the frame data with N samples.
4. The method of claim 1 , wherein the spectral centroid C comprises:
C
=
∑
k
=
1
K
kS
(
k
)
∑
k
=
1
K
S
(
k
)
,
wherein S(k) is a Discrete Fourier Transform (DFT) of frame data s(n) with k frequency components.
5. The method of claim 1 , wherein:
the predetermined threshold TE is a weighted average of the local maxima in a distribution histogram of the short-term energy E; and
the predetermined threshold TC is a weighted average of the local maxima in a distribution histogram of the spectral centroid C.
6. The method of claim 1 , wherein step (b) comprises generating 39-dimensional Mel-Frequency Cepstral Coefficients (MFCCs) from the preprocessed input speech data using overlapping Hamming windows and a hop size.
7. The method of claim 6 , wherein the Hamming window is 25 ms and the hop size of the Hamming windows is 10 ms.
8. The method of claim 1 , wherein the overlapped longer frame length is 10 frames and the hop size of the overlapped longer frames is 3 frames.