US 12,731,603 B2 · Speech features-based single channel voice activity detection method and system for reducing noise from an audio signal

Abstract

The single-channel, Speech Features-Based Voice Activity Detection (SFVAD) system is a robust, low-latency system that generates per-frame speech and noise indications, along with calculating a pair of speech and noise time-frequency masks. The SFVAD system controls an adaptation mechanism for a Beam-Forming system control module and improves the speech quality and noise reduction capabilities of Automatic Speech Recognition applications, such as Virtual Assistance (VA) and Hands-Free (HF) calls, by robustly handling transient noises. The system extracts speech-like patterns from an input audio signal and it is invariant to the power-level of the input audio signal. Noise calculation is controlled by a pair of speech features-based detectors (voiced and unvoiced). A Cepstral-based pitch detector and a Centrum calculation method are used to prevent contamination of the calculated noise by speech content. The SFVAD system robustly handles instant changes of background noise level and has dramatically lower false detection rates.

A computerized method of digital signal processing, comprising:

a) providing an input audio signal comprising speech and noise components;
b) dividing the input audio signal into a sequential series of time frames;
c) identifying a speech segment and/or a noise segment from the input audio signal;
d) extracting a speech features-based pattern from the input audio signal;
e) subtracting the noise segment from the input audio signal;
f) generating an enhanced output audio signal that has reduced noise;
and
r) providing and operating a Single-Channel Voice Activity Detector (SCVAD) module;
wherein the Single-Channel Voice Activity Detector (SCVAD) module comprises:
a Log-Spectral Noise Suppression module;
a Sparsity Detection module;
an Energy-based Speech and Noise Bin-Maps Calculation module;
a Pitch Calculation module;
a Centrum Calculation module;
a Noise Statistics Calculation module;
a Decision Block module;
and
a Z−1 module;
wherein the input audio signal is inputted to the Log-Spectral Noise Suppression module and is also inputted in parallel to the Noise Statistics Calculation module;
wherein a first output from the Log-Spectral Noise Suppression module comprises a Signal-to-Noise Ratio measure, (l,k), that is inputted, in parallel, into four modules that comprise:
the Sparsity Detection Module;
the Energy-based Speech and Noise Bin-Maps Calculation Module;
the Pitch Calculation Module;
and
the Centrum Calculation Module;
wherein a second output from the Sparsity Detection module is inputted into the Decision Block module;
wherein a third output from the Energy-based Speech and Noise Bin-Maps Calculation module is inputted into the Decision Block module;
wherein a fourth output from the Pitch Calculation module is inputted in parallel into both the Decision Block module and the Noise Statistics Calculation module;
wherein a fifth output from the Centrum Calculation module is inputted, in parallel, into both the Decision Block module and the Noise Statistics Calculation module;
wherein a sixth output, <{circumflex over (v)}>(l,k), from the Noise Statistics Calculation module is input to the Z−1 module;
wherein a seventh output from the Z−1 module is fed back into the Log-Spectral Noise Suppression module;
wherein the Decision Block module outputs four parameters:
IS(l), IN(l), MS(l,k), and MN(l,k);
wherein IS(l)=a binary speech indicator that=1 if only speech is present and =0 if not only speech is present;
wherein IN(l)=a binary noise indicator that=1 if only noise is present and =0 if not only noise is present;
wherein MS(l,k)=a Speech Time-Frequency Mask;
wherein MN(l,k)=a Noise Time-Frequency Mask;
wherein “l” represents a selected time frame from the input audio signal, and
wherein “k”′ represents a selected frequency bin from the input audio signal;
and further
g) using the Sparsity Detection module to identify speech utterances by using a sparsity measure that is denoted by S(SNR);
wherein S(SNR) equals a difference between a weighted average of the Signal-to-Noise Ratio (SNR) and an arithmetic average of the SNR in a log-spectrum domain, according to the following equations:
S
⁡
(
d
⁢
B
)
⁢
(
ℓ
)
=
〈
d
⁢
B
〉
w
⁢
(
ℓ
)
-
〈
d
⁢
B
〉
⁢
(
ℓ
)
;
(
eq
.
1
)
〈
d
⁢
B
〉
w
⁢
(
ℓ
)
=
∑
k
=
1
K
⁢
w
⁡
(
ℓ
,
k
)
d
⁢
B
(
ℓ
,
k
)
;
and
(
eq
.
2
)
〈
d
⁢
B
〉
⁢
(
ℓ
)
=
1
K
⁢
∑
k
=
1
K
d
⁢
B
(
ℓ
,
k
)
;
(
eq
.
3
)
wherein S(SNRdB)() denotes a SFVAD sparsity measure of SNRdB;
wherein dB(,k) is weighted by weights w(,k), according to (eq. 2), which accentuate large frequency eigenvalues in the presence of speech;
and
wherein the SFVAD sparsity measure, S(SNRdB)(), is invariant to a power level of the input audio signal, xref(,k);
and
wherein SNRdB(,k) is a suppressed log-spectrum of the input audio signal, xref(,k).

154 more elements

Speech features-based single channel voice activity detection method and system for reducing noise from an audio signal

Life of this patent

Speech features-based single channel voice activity detection method and system for reducing noise from an audio signal

Life of this patent

Abstract

Claims15