This invention relates to signal processing, and more particularly to acquiring and enhancing acoustic signals.
There are numerous applications, such as hands-free mobile communication, telephony, and hands-free speech recognition, where acoustic signals are acquired by a far-field microphone. In such applications, the acquired acoustic signals often include a large amount of noise, e.g., traffic, crowd, radio, TV, wind, or other ambient noise in the environment. The noise interferes with understanding and otherwise processing the acquired acoustic signals, particularly, in speech recognition.
Doppler techniques have been used for a number of applications. U.S. Pat. No. 6,251,077 issued to Mo et al. on Jun. 26, 2001, “Method and apparatus for dynamic noise reduction for Doppler audio output,” describes a method suppressing background noise in spectral Doppler images using adaptive noise-reduction low-pass filters. U.S. Pat. No. 6,773,400 issued to Njemanze on Aug. 10, 2004, “Noninvasive transcranial Doppler ultrasound face and object recognition testing system,” describes a method for measuring a baseline blood flow velocity in cerebral arteries while simultaneously testing the subject with face and object recognition tasks. U.S. Pat. No. 6,773,403 issued to Kim et al. on Aug. 10, 2004, “Ultra-sonic apparatus and method for measuring the velocities of human tissues using the Doppler effects,” describes a method for measuring a velocity of tissue components by sampling echo signals reflected from the human body by generating a frequency distribution of data containing the velocity.
It is desired to use ultra-sonic Doppler techniques for enhancing acoustic signals.
The invention includes a Doppler ultra-sonic signal generator, an acoustic signal detector, e.g., a microphone, and an ultra-sonic signal sensor. A wideband microphone can also be used to acquire both the acoustic and ultra-sonic signals. The ultra-sonic signal has frequencies that are higher than the nominal audio frequency range, e.g., greater than 20 KHz.
During operation, a human speaker speaks while facing the acoustic detector and ultra-sonic sensor. The detector and sensor can be placed on a desktop, on a stand, or be mounted otherwise, e.g., in a car dashboard or on a rearview mirror. The acoustic microphone acquires the acoustic signals generated by the speaker.
Concurrently, the ultrasonic generator transmits a high-frequency acoustic signal towards the speaker's face. The ultra-sonic signal is reflected by the speaker's face, particularly, the area around the mouth, i.e., the lips, tongue, and chin. The reflected Doppler signal is acquired by the ultra-sonic sensor. The frequency of the reflected Doppler signal is modulated by the motion of the speaker's face, tongue, lips and mouth as the acoustic signal is generated.
Therefore, the acoustic and Doppler signals are highly correlated. The acquired ultra-sonic signal is analyzed jointly with the acoustic signal, in order to enhance the acoustic signal. The enhanced acoustic signal can then be further processed for a number of applications, such as in an improved speech recognition system.
System Structure
The system 100 includes a housing 110 that incorporates an acoustic microphone 102, an ultra-sonic transducer 103, and an ultra-sonic sensor 104. The acoustic microphone and ultra-sonic sensor can be replaced by a wideband microphone sensitive over a large range of acoustic frequencies including ultra-sonic Doppler frequencies.
The acoustic microphone 102 acquires a noisy acoustic signal 105 from a source 101, for example, a human speaker. The noisy acoustic signal is converted to an electrical signal 106 representative of the acoustic signal, which is detected 120. It should be noted that the source can be a loudspeaker cone or diaphragm, or some other acoustic source, such a machine with moving parts. In this case, the invention can determine when the machine is operating properly by analyzing jointly the acoustic signals and the reflected Doppler signal modulated according to the moving parts.
An ultra-sonic signal generator 130 produces an ultra-sonic signal 108 for the transducer 103. The ultra-sonic signal is aimed generally at the acoustic source 101, and a reflected Doppler signal 109 is detected separately 140.
When the acoustic signal is produced by moving parts of the acoustic source, e.g., the mouth, lips, and tongue, then the reflected Doppler signal is highly correlated to the corresponding acoustic signal.
Therefore, the invention uses a model that expressed states of ‘clean’ acoustic spectra and corresponding Doppler spectra. The model can then be used to correlate acquired Doppler signals with corresponding clean acoustic signals to enhance the otherwise noisy acoustic signal.
The detected noisy acoustic signal and the Doppler signal are jointly analyzed according to a model 200 to produce 150 an enhanced acoustic signal 151 with a reduced amount of noise. The enhanced acoustic signal 151 can be processed further, for example, to perform a speech recognition 160.
Mixture Model Training
The model has multiple states 220. In each state, there is one spectrum 201 for a ‘clean’ acoustic signal, and one spectrum for a corresponding Doppler signal 202. Such a distribution might be represented mathematically as
where z represents the state, D represents a Doppler spectrum, and S represents the acoustic spectrum, e.g., for a 30 millisecond segment of the acquired signals.
The parameters of the model include a priori probabilities of the states P(z) of the model 200, the state dependent distributions of the Doppler spectra P(D|z), and the state dependent distributions of the acoustic spectra P(S|z). The model is ‘trained’ from a corpus of concurrent ‘clean’ acoustic signals and corresponding Doppler signals. The model can be in the form of a mixture of Gaussian distributions, each distribution having a mean and a variance.
Other models, such as hidden Markov models (HMMs) or Bayesian networks, can also be used.
According to the model, the acoustic signal generation process or source, e.g., the face or a machine, is in a different state at every instant in time. In that state, the source generates a single spectrum for the acoustic signal, and concurrently, a corresponding single spectrum for the Doppler signal that can be correlated to the acoustic signal. The acoustic spectrum 201 is corrupted by the additive noise 203 to generate the noisy acoustic spectrum 204.
Estimating Enhanced Acoustic Signals From Noisy Acoustic Signals
The model for the noise spectrum is used to obtain state-dependent estimates of the enhanced acoustic spectrum 301 from the noisy acoustic signal 105 in an analysis and generating step 310. The noise and Doppler spectra are combined, using the a posteriori probabilities of the states obtained from the Doppler signal, as weights, to generate an estimate for the enhanced acoustic signal 151.
The example procedure can be described as follows. Let Y represent the acoustic spectrum of noisy speech. Let ƒ(S, N) represent a function that expresses the effect of noise N on the acoustic spectrum S of the clean acoustic signal, i.e., Y=ƒ(S, N). The noise is unknown. N is not known.
Ideally, if the clean acoustic spectrum S were also known, then the noise could be estimated through an inverse function N=ƒ1(Y, S). Unfortunately, the clean speech spectrum S is also unknown. However, if the state z that generated the acoustic spectrum S is known, then the noise N can be estimated as N≈ƒ1(Y, μz) where μz is the mean value of the state dependent distributions of the acoustic spectra, P(S|z). However, the state z is also unknown, and cannot be estimated accurately from the acquired acoustic signal because the acoustic signal is noisy. However, the Doppler spectrum D has not been corrupted by noise, and evidence can be derived from the Doppler spectrum to estimate an a posteriori probability of the state z as:
Then, the estimated spectrum for the noise can be obtained as a weighted combination of estimates derived from each of the states, where the weight for the state has been obtained from the Doppler spectrum according to:
N≈P(z|D)ƒ1(Y, μz).
The estimated noise can be used to ‘de-noise’ the spectrum of the acquired acoustic signal through an inverse transformation S≈g(Y, N). This transformation expresses how de-noising can be performed for noisy acoustic signals according to the invention.
It should be understood that the transformation described above is only one example. A variety of transformation functions ƒ(.) and g(.) can be used by the procedure. Similarly other statistical models can also be used represent the various spectra.
A number of applications, such as speech recognition, can operate on the enhanced acoustic signal according to the invention. For example, the invention can be used to de-noise signals in automobiles for hands-free telephony, on cell phones for enhancing transmitted signals when they are used in noisy environments, or to diagnose an operation of machinery and detect potential failures.
Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications may be made within the spirit and scope of the invention. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.
Number | Name | Date | Kind |
---|---|---|---|
4542657 | Barber et al. | Sep 1985 | A |
4864307 | Potage et al. | Sep 1989 | A |
5086775 | Parker et al. | Feb 1992 | A |
5229716 | Demoment et al. | Jul 1993 | A |
5630418 | Lee et al. | May 1997 | A |
6251077 | Mo et al. | Jun 2001 | B1 |
6505057 | Finn et al. | Jan 2003 | B1 |
6773400 | Njemanze | Aug 2004 | B2 |
6773403 | Kim et al. | Aug 2004 | B2 |
7225013 | Geva et al. | May 2007 | B2 |
Number | Date | Country | |
---|---|---|---|
20070165881 A1 | Jul 2007 | US |