Patent Grant
7596494
The present invention relates to speech processing. In particular, the present invention relates to speech enhancement.
In speech recognition, it is common to condition the speech signal to remove noise and portions of the speech signal that are not helpful in decoding the speech into text. For example, it is common to apply a frequency-based transform to the speech signal to reduce certain frequencies in the signal that do not aid in decoding the speech signal. One common frequency-based transform is known as a Mel-Scale transform that reduces pitch harmonics in the speech signal. Mel-Scale transforms are used because the pitch at which someone speaks does not affect the listener's ability to discern what is being said. By removing these harmonics, smaller speech models can be constructed because they do not have to be trained to decode speech at different pitches. Instead, the Mel-scale transform creates pitch-independent models that can be used to decode speech of any pitch.
Speech systems also attempt to enhance the speech signal by removing noise before performing speech recognition. Under some systems, this is done in the time domain by applying a noise filter to the speech signal. In other systems, this enhancement is performed using a two-stage process in which the pitch of the speech is first tracked using a pitch tracker and then the pitch is used to separate the speech signal from the noise. For various reasons, such two-stage processing is undesirable.
A third system for removing noise from a speech signal attempted to identify a clean speech signal in a noisy signal using a probabilistic framework that provided a Minimum Mean Square Error (MMSE) estimate of the clean signal given a noisy signal. This system was designed for speech recognition and as such relied on feature vectors that were appropriate for speech recognition. In particular, this probabilistic system used speech vectors that were produced using the Mel-scale transform.
Although this probabilistic system did not require two-stage processing, it was less than ideal for speech enhancement because the Mel-Scale transform removed information from the signal. Because of this loss of information, it is extremely difficult, if not impossible, to reconstruct a speech signal from the “cleaned” signal that humans can easily understand.
Thus, the current systems for enhancing speech are less than ideal since they either require a two-stage process or make it impossible to reconstruct a clean intelligible speech signal.
A method and apparatus identify a clean speech signal from a noisy speech signal. The noisy speech signal is converted into frequency values in the frequency domain. The parameters of at least one posterior probability of at least one component of a clean signal value are then determined based on the frequency values. This determination is made without applying a frequency-based filter to the frequency values. The parameters of the posterior probability distribution are then used to estimate a set of frequency values for the clean speech signal.
The invention is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems, environments, and/or configurations that may be suitable for use with the invention include, but are not limited to, personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, telephony systems, distributed computing environments that include any of the above systems or devices, and the like.
The invention may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The invention is designed to be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules are located in both local and remote computer storage media including memory storage devices.
With reference to
Computer 110 typically includes a variety of computer readable media. Computer readable media can be any available media that can be accessed by computer 110 and includes both volatile and nonvolatile media, removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared and other wireless media. Combinations of any of the above should also be included within the scope of computer readable media.
The system memory 130 includes computer storage media in the form of volatile and/or nonvolatile memory such as read only memory (ROM) 131 and random access memory (RAM) 132. A basic input/output system 133 (BIOS), containing the basic routines that help to transfer information between elements within computer 110, such as during start-up, is typically stored in ROM 131. RAM 132 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by processing unit 120. By way of example, and not limitation,
The computer 110 may also include other removable/non-removable volatile/nonvolatile computer storage media. By way of example only,
The drives and their associated computer storage media discussed above and illustrated in
A user may enter commands and information into the computer 110 through input devices such as a keyboard 162, a microphone 163, and a pointing device 161, such as a mouse, trackball or touch pad. Other input devices (not shown) may include a joystick, game pad, satellite dish, scanner, or the like. These and other input devices are often connected to the processing unit 120 through a user input interface 160 that is coupled to the system bus, but may be connected by other interface and bus structures, such as a parallel port, game port or a universal serial bus (USB). A monitor 191 or other type of display device is also connected to the system bus 121 via an interface, such as a video interface 190. In addition to the monitor, computers may also include other peripheral output devices such as speakers 197 and printer 196, which may be connected through an output peripheral interface 195.
The computer 110 is operated in a networked environment using logical connections to one or more remote computers, such as a remote computer 180. The remote computer 180 may be a personal computer, a hand-held device, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to the computer 110. The logical connections depicted in
When used in a LAN networking environment, the computer 110 is connected to the LAN 171 through a network interface or adapter 170. When used in a WAN networking environment, the computer 110 typically includes a modem 172 or other means for establishing communications over the WAN 173, such as the Internet. The modem 172, which may be internal or external, may be connected to the system bus 121 via the user input interface 160, or other appropriate mechanism. In a networked environment, program modules depicted relative to the computer 110, or portions thereof, may be stored in the remote memory storage device. By way of example, and not limitation,
Memory 204 is implemented as non-volatile electronic memory such as random access memory (RAM) with a battery back-up module (not shown) such that information stored in memory 204 is not lost when the general power to mobile device 200 is shut down. A portion of memory 204 is preferably allocated as addressable memory for program execution, while another portion of memory 204 is preferably used for storage, such as to simulate storage on a disk drive.
Memory 204 includes an operating system 212, application programs 214 as well as an object store 216. During operation, operating system 212 is preferably executed by processor 202 from memory 204. Operating system 212, in one preferred embodiment, is a WINDOWS® CE brand operating system commercially available from Microsoft Corporation. Operating system 212 is preferably designed for mobile devices, and implements database features that can be utilized by applications 214 through a set of exposed application programming interfaces and methods. The objects in object store 216 are maintained by applications 214 and operating system 212, at least partially in response to calls to the exposed application programming interfaces and methods.
Communication interface 208 represents numerous devices and technologies that allow mobile device 200 to send and receive information. The devices include wired and wireless modems, satellite receivers and broadcast tuners to name a few. Mobile device 200 can also be directly connected to a computer to exchange data therewith. In such cases, communication interface 208 can be an infrared transceiver or a serial or parallel communication connection, all of which are capable of transmitting streaming information.
Input/output components 206 include a variety of input devices such as a touch-sensitive screen, buttons, rollers, and a microphone as well as a variety of output devices including an audio generator, a vibrating device, and a display. The devices listed above are by way of example and need not all be present on mobile device 200. In addition, other input/output devices may be attached to or found with mobile device 200 within the scope of the present invention.
The present invention provides a method and apparatus for reconstructing a speech signal using high resolution speech vectors.
At step 400, a noisy analog signal 300 is converted into a sequence of digital values that are grouped into frames by a frame constructor 302. Under one embodiment, the frames are constructed by applying analysis windows to the digital values where each analysis window is a 25 millisecond hamming window, and the centers of the windows are spaced 10 milliseconds apart.
At step 402, a frame of the digital speech signal is provided to a Fast Fourier Transform 304 to compute the phase and magnitude of a set of frequencies found in the frame. Under one embodiment, Fast Fourier Transform 304 produces noisy magnitudes 306 and phases 308 for 128 frequencies in each frame. The phases 308 for the frequencies are stored for later use. A log function 310 is applied to magnitudes 306 at step 408 to compute the logarithm of each magnitude.
At step 410, the logarithm of each magnitude is provided to a finite impulse response (FIR) filter 312, which filters each magnitude over time. Under one embodiment, the FIR filter uses three consecutive frames for filtering using filter parameters of (0.25 0.5 0.25). This smoothes the log magnitudes and reduces spurious errors.
The filtered log magnitudes are provided as a vector of magnitude values to a posterior calculator 314, which computes a posterior probability for the vector at step 410. The posterior probability provides the probability of a clean speech log magnitude vector given the noisy speech filtered log magnitude vector. Under one embodiment, a mixture model is used consisting of a mixture of different posterior components, each having a mean and variance. Under one specific embodiment, a mixture model consisting of 512 male speaker mixture components and 512 female speaker mixture components is used. One technique for computing the posterior probabilities is discussed further below in connection with
At step 414 the posterior probability is used to compute an estimate of the clean log magnitude spectrum using an estimator 316. Under one embodiment, the estimate of the clean log magnitude spectrum is a weighted average of the minimum mean square error estimates calculated from each of the mixture components of the posterior probability.
The estimated clean signal log magnitude values are exponentiated at step 416 by an exponent function 318 to produce estimates of the clean magnitudes 320. At step 418, an inverse Fast Fourier Transform 322 is applied to the clean magnitudes 320 using the stored phases 308 taken from the noisy signal at step 402 above. The inverse Fast Fourier Transform results in a frame of time domain digital values for the frame.
At step 420 an overlap and add unit 326 is used to overlap and add the frames of digital values produced by the inverse Fast Fourier Transform to produce a clean digital signal 328. Under one embodiment, this is done using synthesis windows that are designed to provide perfect reconstruction when the analyzed signal is perfect and to reduce edge effects. Under one particular embodiment, when an analysis window of a(s) is used, the synthesis window, b(s) is defined as:
where τ is the time period between the beginning of successive analysis windows and the summation is taken over the number of windows.
The output clean digital signal 328 can then be written to output audio hardware so that it is perceptible to users or stored at step 422.
As shown above, the present invention does not apply a frequency-based transform to the noisy log-magnitude values before determining the posterior probability. A frequency-based transform is one in which the level of filtering applied to a frequency is based on the identity of the frequency or the magnitudes of the frequencies are scaled and combined to form fewer parameters. (Note that the FIR filter in
By retaining all of the frequencies in the feature vector, the present invention provides a better posterior calculation, and thus a better estimate for the clean speech frequencies. In addition, because the number of frequency bins has not been reduced, the reconstructed signal is more intelligible, since information was not lost through a Mel-Scale transform.
A process for identifying the posterior probability p(nxc|y) of noise channel distortion, c, and clean signal, x, given a noisy signal y, is shown in
To generate the means and variances of the prior probability, the process of one embodiment of the present invention first generates a mixture of Gaussians that describes the distribution of a set of training noise feature vectors, a second mixture of Gaussians that describes a distribution of a set of training channel distortion feature vectors, and a third mixture of Gaussians that describes a distribution of a set of training clean signal feature vectors. The mixture components can be formed by grouping training feature vectors using a maximum likelihood training technique or by grouping training feature vectors that represent a temporal section of a signal together. Those skilled in the art will recognize that other techniques for grouping the feature vectors into mixture components may be used and that the two techniques listed above are only provided as examples. Under one embodiment, one mixture component is used for noise, one mixture component is used for channel distortion, and 128 mixture components are used for clean speech.
After the training feature vectors have been grouped into their respective mixture components, the mean and variance of the feature vectors within each component is determined. In an embodiment in which maximum likelihood training is used to group the feature vectors, the means and variances are provided as by-products of grouping the feature vectors into the mixture components.
After the means and variances have been determined for the mixture components of the noise feature vectors, clean signal feature vectors, and channel feature vectors, these mixture components are combined to form a mixture of Gaussians that describes the total prior probability. Using one technique, the mixture of Gaussians for the total prior probability will be formed at the intersection of the mixture components of the noise feature vectors, clean signal feature vectors, and channel distortion feature vectors.
The variances of the mixture components of the observation probability are determined using a closed form expression of the form:
where α is estimated from the training data.
Under other embodiments, these variances are formed using a training clean signal, a training noise signal, and a set of training channel distortion vectors that represent the channel distortion that will be applied to the clean signal and noise signal.
The training clean signal and the training noise signal are separately converted into sequences of feature vectors. These feature vectors, together with the channel distortion feature vectors are then applied to an equation that approximates the relationship between observed noisy vectors and clean signal vectors, noise vectors, and channel distortion vectors. Under one embodiment, this equation is of the form:
y≈c+x+(ln(1+e([n−c−x]))) EQ. 3
where y is an observed noisy feature vector, c is a channel distortion feature vector, x is a clean signal feature vector, and n is a noise feature vector. In equation 3:
where nj, cj, and xj are the jth elements in the noise feature vector, channel feature vector, and clean signal feature vector, respectively.
Under one embodiment, the training clean signal feature vectors, training noise feature vectors, and channel distortion feature vectors used to determine the mixture components of the prior probability are reused in equation 3 to produce calculated noisy feature vectors. Thus, each mixture component of the prior probability produces its own set of calculated noisy feature vectors.
The training clean signal is also allowed to pass through a training channel before being combined with the training noise signal. The resulting analog signal is then converted into feature vectors to produce a sequence of observed noisy feature vectors. The observed noisy feature vectors are aligned with their respective calculated noisy feature vectors so that the observed values can be compared to the calculated values.
For each mixture component in the prior probability, the average difference between the calculated noisy feature vectors associated with that mixture component and the observed noisy feature vectors is determined. This average value is used as the variance for the corresponding mixture component of the observation probability. Thus, the calculated noisy feature vector produced from the third mixture component of the prior probability would be used to produce a variance for the third mixture component of the observation probability. At the end of step 500 a variance has been calculated for each mixture component of the observation probability.
After the parameters of the mixture components of the prior probability and the observation probability have been determined, the process of
Due to the non-linear relationship in Equation 3, the true posterior is non-Gaussian. However, under one embodiment of the invention, the posterior is approximated as a Gaussians. In order to make this approximation, a linear approximation of Equation 3 must be made. This is done using a first order Taylor series expansion of:
y≅g(zo)+g′(zo)(z−zo) EQ. 5
where z and zo are stacked vectors representing a combination of a noise vector, channel vector and clean signal vector such that
z=[xTnTcT] EQ. 6
zo=[xoTnoTcoT] EQ. 7
and where
g(zo)=xo+co+1n(1+e[n
and g′(zo) is the derivative of g(zo) determined at expansion point zo.
Using the Taylor series expansion, the variance and mean and variance of the posterior probability can be calculated iteratively using:
η=ηp+Φ(Σ−1(μ−ηp)+g′(ηp)TΨ−1(y−g(ηp))) EQ. 9
Φ=(Σ−1+g′(ηp)TΨ−1g′(ηp))−1 EQ. 10
where η is the newly calculated mean for the posterior probability of the current mixture, ηp is the mean for the posterior probability determined in a previous iteration, Σ−1 is the inverse of the covariance matrix for this mixture component of the prior probability, μ is the mean for this mixture component of the prior probability, Ψ is the variance of this mixture component of the observation probability, Φ is the variance of the posterior probability for this mixture component, g(ηp) is the right-hand side of equation 8 evaluated with the expansion point set equal to the mean of the previous iteration, g′(ηp) is the matrix derivative of equation 8 calculated at the mean of the previous iteration, and y is the observed feature vector.
In equation 9, μ, η and ηp are M-by-1 matrices where M is three times the number of elements in each feature vector. In particular, μ, η and ηp are described by vectors having the form:
Using this definition for μ, η and ηp, and using ηp as the expansion point zo, Equation 8 above can be described as:
where the designations in equation 12 indicate the spans of rows which form the feature vectors for those elements.
In equations 9 and 10, the derivative g′(ηp) is a matrix of order
where the element of row i, column j is defined as:
where the expression on the right side of equation 13 is a partial derivative of the equation that describes the ith element of g(ηp) relative to the jth element of the ηp matrix. Thus, if the jth element of the ηp matrix is the fifth element of the noise feature vector, n5, the partial derivative will be taken relative to n5.
The iterative process for determining the means and variance of the posterior probability is shown in steps 504, 506, 508, 510 and 512 of
If more iterations are to be performed, the current mean η is set as the past mean ηp at step 512 so that the current mean is used as the expansion point in the next iteration. The process then returns to step 506. Steps 506, 508, 510 and 512 are then repeated until the desired number of iterations has been performed.
After the mean and variance for the first mixture component of the posterior probability has been determined, the process of
Once a mean and variance has been determined for each mixture component of the posterior probability, the process of
where S is the number of mixture components, ρs is the weight for mixture component s,
is the feature vector for the mean of the posterior probability of the clean signal, and xpost is the weighted average value of the clean signal feature vector given the observed noisy feature vector.
The weight for each mixture component, ρs is calculated as:
where the dominator of equation 15 normalizes the weights by dividing each weight by the sum of all other weights for the mixture components. In equation 15, πs is a weight associated with the mixture components of the prior probability and is determined as:
πs=πsx·πsn·πsc EQ. 16
where πsx, πsn, and πsc are mixture component weights for the prior clean signal, prior noise, and prior channel distortion, respectively. These weights are determined as part of the calculation of the mean and variance for the prior probability.
In equation 15, Gs is a function that affects the weighting of a mixture component based on the shape of the prior probability and posterior probability, as well as the similarity between the selected mean for the posterior probability and the observed noisy vector and the similarity between the selected mean and the mean of the prior probability. Under one embodiment, the expression for Gs is:
where ln|2πΣs| involves taking the natural log of the determinant of 2π times the covariance of the prior probability, ln|2πΦs| involves taking the natural log of the determinant of 2π times the covariance matrix of the posterior probability.
In other embodiments, the clean signal vector is estimated as:
Those skilled in the art will recognize that there are other ways of using the mixture approximation to the posterior to obtain statistics. For example, the means of the mixture component with largest ρ can be selected. Or, the entire mixture distribution can be used as input to a recognizer.
Although a particular method for determining the posterior probability is discussed above, those skilled in the art will recognize that any technique for identifying the posterior probability may be used with the present invention.
Although the present invention has been described with reference to particular embodiments, workers skilled in the art will recognize that changes may be made in form and detail without departing from the spirit and scope of the invention.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5148489 | Erell et al. | Sep 1992 | A |
| 5924065 | Eberman et al. | Jul 1999 | A |
| 6026359 | Yamaguchi et al. | Feb 2000 | A |
| 6067517 | Bahl et al. | May 2000 | A |
| 6188976 | Ramaswamy et al. | Feb 2001 | B1 |
| 6195632 | Pearson | Feb 2001 | B1 |
| 6202047 | Ephraim et al. | Mar 2001 | B1 |
| 6633842 | Gong | Oct 2003 | B1 |
| Number | Date | Country | |
|---|---|---|---|
| 20050114117 A1 | May 2005 | US |
