ATIS 0100005-2006 Auditory Non-Intrusive Quality Estimation Plus (ANIQUE+) Perceptual Model for Non-Intrusive Estimation of Narrowband Speech Quality (Includes Access to Additional.pdf

1、 Access to Additional Content for ATIS-0100005, Dated: December 2006 (R2011) (Click here to view the publication) This Page is not part of the original publication This page has been added by IHS as a convenience to the user in order to provide access to additional content as authorized by the Copyr

2、ight holder of this document Click the link(s) below to access the content and use normal procedures for downloading or opening the files. ATIS-0100005 Information contained in the above is the property of the Copyright holder and all Notice of Disclaimer their existence does not in any respect prec

3、lude anyone, whether he has approved the standards or not, from manufacturing, marketing, purchasing, or using products, processes, or procedures not conforming to the standards. The American National Standards Institute does not develop standards and will in no circumstances give an interpretation

4、of any American National Standard. Moreover, no person shall have the right or authority to issue an interpretation of an American National Standard in the name of the American National Standards Institute. Requests for interpretations should be addressed to the secretariat or sponsor whose name app

5、ears on the title page of this standard. CAUTION NOTICE: This American National Standard may be revised or withdrawn at any time. The procedures of the American National Standards Institute require that action be taken periodically to reaffirm, revise, or withdraw this standard. Purchasers of Americ

6、an National Standards may receive current information on all standards by calling or writing the American National Standards Institute. Notice of Disclaimer n), which is the temporal envelope for the k-th critical band at the m-th frame. The modulation spectrum for each critical band is then estimat

7、ed by the magnitude of Fourier transform as: k(m,f) = |Fk(m;n)| where f denotes modulation frequency and Fx is the Fourier transform of x. The modulation spectrum is grouped into 7 bands by a modulation filterbank Wmod(i,f), i = 1, 2, , 7, which is implemented and applied in modulation frequency dom

8、ain, and one can obtain modulation band power of the k-th critical band at the m-th time frame as: = dffiWfmimkk),(),(),(2mod2. For discrete-time signals, the modulation spectrum is obtained by 2048-point FFT and the modulation band power can be approximated as . =102302mod2),(),(),(jkkjiWjmimFigure

9、 3 shows the frequency response of modulation filterbank Wmod(i,f). The quality factor of each filter is set 2 which is defined as the center frequency divided by the bandwidth of filter. 7 ATIS-0100005.2006 2 4 8 16 32 64 128 25600.20.40.60.81Modulation Frequency HzGainFigure 3 - Frequency response

10、 of modulation filterbank The average articulation power is defined as: =ALikAAkimLm1,),(1)( and reflects the amount of signal components relevant to natural human speech. In order to cover the frequency range of 2 30 Hz, corresponding to the movement speed of human articulation system, LAis set to

11、4. On the contrary, the average nonarticulation power represents the amount of perceptually annoying distortions produced at the rates beyond the speed of human articulation systems, and is defined as: +=)(1,),()(1)(kLLikANNkNAimLkLm Where: =221971814,6130,5)(kkkkLNspecifies the last modulation filt

12、er index to be considered in estimating the nonarticulation power. 8.3 Feature Extraction The logarithm of the DC-value of modulation spectrum is subtracted from the logarithms of average articulation power and the average nonarticulation power to represent normalized values as: 8 ATIS-0100005.2006

13、)1)0,(log( -1)(m)(log)(kAk,+= mmAkAnd: )1)0,(log( -1)(m)(log)(kNk,+= mmNk. Also, )1)0,(log(max)1)0,(log()0,(k+= mmmjjkis the normalized DC-value of modulation log-spectrum. The input to frame distortion model at the m-th frame is a 69-dimensional feature vector: (m) = A(m); N(m); (m,0) where A(m), N

14、(m), and (m,0) are row vector representations of k,A(m), k,N(m), and k(m,0), respectively. 9 DETECTION OF ACTIVE SPEECH AND AUDIBLE BACKGROUND NOISE The voice activity detection (VAD) is estimated every 16 msec based on frame power, its time-derivative, and estimated adaptive background noise power.

15、 The frame power calculated every 16 msec, P16ms(m), is defined as: +=1)();(log10)(255023221016nmsmsnHammnmsmPwhere s(m;n) is the m-th, 32 msec frame of s(n), advanced in 16 msec steps and is the 32 msec-long Hamming window: )(32nHammmsfor 0 n N-1 ),1/(2cos(46.054.0)(32= NnnHammmswhere N = 256. The

16、time-derivative of P16ms(m) is calculated as: 2)1()1()(161616+= mPmPmPmsmsms. Figure 4 shows the overview of the VAD procedure. For every frame, a threshold is adaptively determined based on the estimated background noise level and is used to obtain VAD profile - whether the current frame belongs to

17、 active speech or background noise. The VAD profile is then post-processed to obtain the final VAD profile with samples spaced every 16 msec. In order to use the VAD profile in the frame distortion estimation (Clause 10), the estimated VAD profile with 16 msec sampling is then transformed to a final

18、 estimated VAD profile with 64 msec sampling: =.,0,1)(otherwisespeechactiveformISAmong the frames determined as background noise, those of which the value of frame envelope (Penv(m) defined in Section 10) exceeds a threshold ( = 82) are considered as audible background noise, to obtain the profile f

19、or audible background noise: 9 ATIS-0100005.2006 =.,0,1)(otherwisenoisebackgroundaudibleformIBFrame PowerThreshold EstimationTime-derivative of Frame PowerSpeech / Background Noise DecisionBackground Noise Power Estimation for Next FrameVAD Profile (16 ms)VAD Profile IS(64 ms)Post-processingAudible

20、Background Noise Profile IB(64 ms)Transformation to 64 ms IntervalFigure 4 - Detection of active speech and audible background noise 10 FRAME DISTORTION MODEL 10.1 Overview In the frame distortion model, the perceptual distortion of individual speech frames is estimated every 64 msec and the overall

21、 frame distortion DFis modeled as: DF= DS+ DB . Here, DSis the distortion in speech obtained by accumulating frame distortions over active speech frames (IS(m) = 1 in Clause 9) and then normalizing by the total number of active speech frames TSas: =SmSSmTD )(1 where (m) is the output of the frame di

22、stortion model ranging from 0 to 1 at the m-th frame in 64 msec interval. DBis the distortion in background noise and is estimated for audible background noise frames (IB(m) = 1 in Clause 9) as: ()+=BmFthenvFBBmPmPTD )()(1 where TBis the number of frames determined as audible background noise, Pth(=

23、 82.0) is the threshold to determine whether the background noise is audible enough, and Penv(m) is the frame envelope defined as 10 ATIS-0100005.2006 . =cbNkkenvmmP1210)0,(log10)(Two parameters F( = 0.0064 ) and F( = 0.014696 ) are weighting factors for the frame envelope. Figure 5 illustrates how

24、the frame distortion is estimated from a speech signal. The feature vector (m) at the m-th frame is fed into the frame distortion model as an input to produce the frame distortion (m). If the current frame is classified as active speech frame, the frame distortion value is accumulated to produce the

25、 distortion for active speech DS. If the current frame is classified as audible background noise, the frame distortion is weighted through the frame envelope and accumulated to yield the background distortion DB. These two distortion values, DSand DB, are added to obtain the overall frame distortion

26、 DF. 11 ATIS-0100005.2006 Figure 5 - Frame distortion estimation model 10.2 Multi-Layer Perceptron Model The frame distortion model, F, is the multi-layer perceptron (MLP) with one hidden layer as shown in Figure 6. 12 ATIS-0100005.2006 Figure 6 - Multi-layer perceptron for frame distortion model Th

27、e MLP consists of 69 input neurons, 120 hidden neurons, one output neuron, and synaptic weights connecting between layers. It can learn the mapping function between input feature vectors and the corresponding perceptual distortions derived from subjective MOS value. The output of MLP is expressed as

28、: =jkkjkjmwgWgmy )()( Here, k(m) is the k-th element of input feature vector (m), wjkand Wjare synaptic weights for the input and hidden layer, respectively, and g(x) is the nonlinear sigmoid function defined as: 1)exp(12)( +=xxgMLPwhere MLP(= 0.3) is the slope of sigmoid function. As the range of o

29、utput of MLP, y(m), is (-1, 1), the estimated frame distortion can be obtained: 8.1/)9.0)()( += mym Where: . 3This document is available from the Alliance for Telecommunications Industry Solutions (ATIS), 1200 G Street N.W., Suite 500, Washington, DC 20005. 4This document is available from the European Telecommunications Standards Institute (ETSI). ATIS-0100005.2006 20 Annex B (normative) B ANSI-C REFERENCE IMPLEMENTATION OF ANIQUE+ This Annex has been formatted as a separate folder containing the C-source code of the ANIQUE+ algorithm, and electronically packaged with this standard.