Voice activity detector
View Patent ↗A voice activity detector indicates the presence of speech within a signal. The detector may determine whether the signal includes speech by calculating a variance of a signal-to-noise ratio across a plurality of portions of a signal, calculating a value based on the variance of the signal-to-noise ratio, performing a comparison between the value and a threshold, and identifying whether the signal contains speech based on the comparison between the value and the threshold.
1. A voice activity detection method, comprising:
calculating, by a processor, a variance of a signal-to-noise ratio across a plurality of portions of a signal;
calculating, by the processor, a value based on the variance of the signal-to-noise ratio;
performing, by the processor, a comparison between the value and a threshold; and
identifying, by the processor, whether the signal contains speech based on the comparison between the value and the threshold.
2. The method of claim 1 , where the step of calculating the value comprises combining a plurality of signal-to-noise ratio variance measurements calculated for a plurality of wide bands of the signal to derive the value.
3. The method of claim 2 , where the step of identifying whether the signal contains speech comprises identifying that the signal contains speech in response to a determination that the value exceeds the threshold, and identifying that the signal does not contain speech in response to a determination that the value is less than the threshold.
4. The method of claim 2 , where the step of combining the plurality of signal-to-noise ratio variance measurements comprises applying a weighting function that weights the plurality of signal-to-noise ratio variance measurements and combines them into a single value.
5. The method of claim 2 , further comprising temporally smoothing the value before comparing the value to the threshold.
6. The method of claim 1 , further comprising dividing the signal into a wide band structure for noise estimation, and storing the wide band structure of the signal in computer memory for use by a voice activity detector.
7. The method of claim 1 , where the step of calculating the value comprises deriving a temporally smoothed, weighted sum of a plurality of signal-to-noise ratio variance measurements of a plurality of wide bands of the signal.
8. The method of claim 1 , where the step of calculating the variance of the signal-to-noise ratio comprises calculating an average difference between a signal measurement at each bin of a portion of the signal and a noise estimate at each bin of the portion of the signal.
9. The method of claim 1 , where the signal is divided into multiple wide bands and multiple bins within the wide bands, and where the step of calculating the variance of the signal-to-noise ratio comprises calculating the variance according to:
V
j
=
∑
0
N
-
1
(
S
i
-
D
i
)
2
N
-
(
∑
0
N
-
1
S
i
-
∑
0
N
-
1
D
i
N
)
2
where V j is the variance of the signal-to-noise ratio, S i is an estimate of the signal at bin “i” within wide band “j,” and D i is an estimate of a noise at bin “i” within wide band “j.”
10. A voice activity detection method, comprising:
dividing, by a processor, a signal into a plurality of wide bands;
dividing, by the processor, each of the wide bands into a plurality of bins;
determining, by the processor, a noise estimate for each of the wide bands;
calculating, by the processor for each of the wide bands, a variance of a signal-to-noise ratio across the bins of each of the wide bands based on the signal and the noise estimate for each of the wide bands;
combining, by the processor, the variances calculated for each of the wide bands to derive a value;
performing, by the processor, a comparison between the value and a threshold; and
identifying, by the processor, whether the signal contains speech based on the comparison between the value and the threshold.
11. The method of claim 10 , where the step of identifying whether the signal contains speech comprises identifying that the signal contains speech when the value exceeds the threshold and identifying that the signal does not contain speech when the value is less than the threshold.
12. The method of claim 10 , where the step of calculating the variance of the signal-to-noise ratio comprises calculating the variance according to:
V
j
=
∑
0
N
-
1
(
S
i
-
D
i
)
2
N
-
(
∑
0
N
-
1
S
i
-
∑
0
N
-
1
D
i
N
)
2
where V j is the variance of the signal-to-noise ratio, S i is an estimate of the signal at bin “i” within wide band “j,” and D i is an estimate of a noise at bin “i” within wide band “j.”
13. A noise detection system, comprising:
a computer memory that stores a measurement of a variance of a signal-to-noise ratio across a plurality of portions of a signal; and
a processor coupled with the computer memory;
where the processor is configured to access the measurement of the variance of a signal-to-noise ratio from the computer memory;
where the processor is configured to calculate a value based on the variance of the signal-to-noise ratio;
where the processor is configured to perform a comparison between the value and a threshold; and
where the processor is configured to identify whether the signal contains speech based on the comparison between the value and the threshold.
14. The system of claim 13 , where the processor is configured to combine a plurality of signal-to-noise ratio variance measurements calculated for a plurality of wide bands of the signal to derive the value.
15. The system of claim 14 , where the processor is configured to identify that the signal contains speech in response to a determination that the value exceeds the threshold, and identify that the signal does not contain speech in response to a determination that the value is less than the threshold.
16. The system of claim 14 , where the processor is configured to apply a weighting function that weights the plurality of signal-to-noise ratio variance measurements and combines them into a single value.
17. The system of claim 14 , where the processor is configured to temporally smooth the value before comparing the value to the threshold.
18. The system of claim 13 , where the processor is configured to derive the value as a temporally smoothed, weighted sum of a plurality of signal-to-noise ratio variance measurements of a plurality of wide bands of the signal.
19. The system of claim 13 , where the processor is configured to calculate an average difference between a signal measurement at each bin of a portion of the signal and a noise estimate at each bin of the portion of the signal.
20. The system of claim 13 , where the signal is divided into multiple wide bands and multiple bins within the wide bands, and where the processor is configured calculate the variance according to:
V
j
=
∑
0
N
-
1
(
S
i
-
D
i
)
2
N
-
(
∑
0
N
-
1
S
i
-
∑
0
N
-
1
D
i
N
)
2
where V j is the variance of the signal-to-noise ratio, S i is an estimate of the signal at bin “i” within wide band “j,” and D i is an estimate of a noise at bin “i” within wide band “j.”