Ultrasonic doppler sensor for speaker recognition
View Patent ↗A method and system recognizes an unknown speaker by directing an ultrasonic signal at a face of the unknown speaker. A Doppler signal of the ultrasonic signal is acquired after reflection by the face, and Doppler features are extracted from the reflected Doppler signal. The Doppler features are classified using Doppler models storing the Doppler features and identities of known speakers to recognize and identify the unknown speaker.
1. A method for recognizing an unknown speaker, comprising the steps of:
directing an ultrasonic signal at a face of an unknown speaker;
acquiring a Doppler signal of the ultrasonic signal after reflection by the face;
extracting Doppler features from the reflected Doppler signal; and
classifying the Doppler features using Doppler models storing the Doppler features and identities of known speakers to recognize and identify the unknown speaker, wherein there is one model for each known speaker.
2. The method of claim 1 , further comprising:
acquiring speech signals from the unknown speaker, while acquiring the Doppler signal;
extracting audio features from the speech signal; and
classifying the Doppler features and the audio features using the Doppler models and audio models storing the speech features and the identities of the known speakers to recognize and identify the unknown speaker.
3. The method of claim 1 , further comprising:
acquiring visual signals from the unknown speaker, while acquiring the Doppler signal;
extracting visual features from the visual signal; and
classifying the visual features and the Doppler features using the Doppler models and visual models storing the visual features and the identities of the known speakers to recognize and identify the unknown speaker.
4. The method of claim 1 , further comprising:
acquiring speech signals from the unknown speaker, while acquiring the Doppler signal;
extracting audio features from the speech signal;
acquiring visual signals from the unknown speaker, while acquiring the Doppler signal and the speech signal;
extracting visual features from the visual signal; and
classifying the visual features and the Doppler features and the speech features using the Doppler models and visual models storing the visual features and speech models storing the speech features and the identities of the known speakers to recognize and identify the unknown speaker.
5. The method of claim 1 , wherein the ultrasonic signal is a 40 kHz tone.
6. The method of claim 2 , wherein means for acquiring the Doppler and speech signals are collocated.
7. The method of claim 1 , further comprising:
band-pass sampling the Doppler signal to a bandwidth of 4 KHz.
8. The method of claim 6 , wherein the ultrasonic signal has a frequency f and the face is moving with a velocity v relative to the means for acquiring, and a frequency {circumflex over (f)} of the Doppler signal is
f
^
=
υ
s
+
υ
υ
s
-
υ
f
(
1
)
where v s is a speed of sound in a particular medium.
9. The method of claim 8 , wherein the face includes multiple articulators, and each articulator has a particular velocity to produce an associated Doppler signal with a corresponding frequency.
10. The method of claim 9 , wherein Doppler signal over time is expressed as
d
(
t
)
=
∑
i
=
1
N
a
i
(
t
)
cos
(
2
π
f
i
(
t
)
+
ϕ
i
)
+
Ψ
speaker
,
(
2
)
where f i is a frequency of the Doppler signal for an i th articulator, which is dependent on a velocity v i of the i t articulator, and the frequency of the ultrasonic signal is f c , and a i (t) is a time-varying reflection coefficient that is related to a distance between the i th articulator and the means for acquiring, and an articulator-specific phase correction term is φ i , and Ψ speaker is speaker dependent.
11. The method of claim 1 , wherein the classifying uses a Bayesian classifier, and wherein each model includes a distribution of the Doppler features of the known speakers.
12. The method of claim 11 , wherein the distribution for known speaker w is
P
(
D
|
w
)
=
∑
i
c
w
,
i
d
𝒩
(
D
;
μ
w
,
i
d
,
R
w
,
i
d
)
where D represent a random feature vectors derived from the Doppler signal, P(D|w) represent the distribution of Doppler feature vectors for speaker w, N(X; μ, R) represents a value of a multivariate Gaussian mixture with mean μ and covariance R at a point X, and μ d w,i , R d w,i and c d w,i represent a mean, covariance matrix and mixture weights for an i th Gaussian mixture in the distribution of the Doppler features for the speaker w.
13. The method of claim 12 , wherein the unknown speaker is recognized as known speaker ŵ according a rule:
w
^
=
ar
arg
max
w
P
(
w
)
∏
D
∈
D
P
(
D
|
w
)
,
where P(w) represents an a priori probability of the speaker w.
14. The method of claim 2 , wherein contributions of the Doppler and speech features during the classifying are weighted.
15. The methods of claims 1 , further comprising:
rejecting the unknown speaker using a generic model.
16. A system for recognizing an unknown speaker, comprising:
a transmitter configured to direct an ultrasonic signal at a face of a speaker;
a receiver configured to acquire a Doppler signal after the ultrasonic signal is reflected by the face;
means for extracting Doppler features from the reflected Doppler signal; and
means for classifying the Doppler features using Doppler models storing the Doppler features and identities of known speakers to recognize and identify the unknown speaker, wherein there is one model for each known speaker.