The invention relates to speech quality estimation, and more particularly to a method, a computer program, a computer program product, and a computer for speech quality estimation.
Bandwidth limitations and signal presentation level variations affect the overall perception of speech quality. Presentation level is the active speech level at the listener side. How to measure active speech level is described in [1] ITU-T Rec. P. 56 (March 1993) Objective measurement of Active Speech Level.
If the bandwidth and the presentation level variations are the only source of degradation, they can be related in a simple way to speech quality; the signals with larger bandwidth and higher presentation level have higher quality and vice versa. However, in the case of typical coding artifacts, this relation becomes highly non-linear, and limiting the signal bandwidth and/or decreasing presentation level might lead to quality improvement. This effect is difficult to capture by the conventional quality assessment schemes, such as those disclosed in the following documents [2]-[6] below:
[2] ITU-T Rec. P.862 (February 2001), Perceptual evaluation of speech quality (PESQ), an objective method for end-to-end speech quality assessment in narrow-band telephone networks and speech codecs;
[3] ITU-T Rec. P.862.2 (November 2005), Wideband extension to Recommendation P.862 for the assessment of wideband telephone networks and speech codecs;
[4] ANSI T1.518-1998 (R2003), Objective Measurement of Telephone Band Speech Quality Using Measuring Normalizing Blocks;
[5] ITU-T P. 563 (May 2004), Single ended method for objective speech quality assessment in narrow-band telephony applications; and
[6] ITU-R Rec. BS.1387-1 (November 2001), Method for objective measurements of perceived audio quality.
Presentation level is related to the signal loudness, typically measured according to ITU-T Rec. P.56 speech level meter described in [1]. An example of a signal at different presentation levels is shown in
Signal bandwidth is the range of frequencies beyond which the frequency function is close to zero (e.g. 10-20 dB below max frequency value). Example of a super-wideband signal (50-14000 Hz), processed with NB (narrowband) IRS (Intermediate Reference System) filter is given in
An object of the invention is to improve speech quality estimation, i.e. improve the assessment of speech quality of a speech signal.
The invention relates to a method performed by a computer for speech quality estimation. The method comprises the steps of:
Q
COD+ω1·BW+ωwPL, and
Hereby bandwidth limitations and presentation level variations are taken into account. The invention presents a scheme that can capture the non-linear relation between a coding noise, a bandwidth variation, and a presentation level variation, but is still simple and thus generalizes better with unknown data. In this way the effects of BW and PL can be incorporated in a more general quality assessment scheme, without causing problems related to data overfitting.
In one embodiment of the method, the step of extracting ω1 and ω2 is performed by calculating ωi=
∥QCOD−γi∥α
where i={1, 2} and wherein γ and α are trained or empirically determined coefficients.
In one embodiment of the method, the step of extracting ω1 and ω2 is performed by calculating ωi=
−∥Q
COD−γi∥β
where i={1, 2} and wherein γ and β are trained or empirically determined coefficients.
In one embodiment of the method, the step of extracting ω1 and ω2 is performed by calculating ω1 and ω2 according to
where i={1, 2} and γ, α and β are trained or empirically determined coefficients.
QCOD may be determined by extracting QCOD from
wherein N is a number of frames or blocks in the speech signal and W is a number of frequency bands wherein the N and the W are related to a codec bit rate with n being a time frame, frame index or frame counter value and f being a frequency counter or band index value, and P represents power spectrum of the speech signal.
Q may in one embodiment of the method be used to
The invention also relates to a computer for speech quality estimation. The computer is adapted to be connected to a communications network and comprises:
The computer may comprise a speech quality estimation unit configured to use Q to estimate a speech quality of the speech signal.
The computer may comprise an input unit for receiving an original signal and a processed signal of the original signal.
The extracting unit of the computer may be configured to extract ω1 and ω2 by calculating ωi=
∥QCOD−γi∥α
where i={1, 2} and wherein γ and α are trained or empirically determined coefficients.
The extracting unit of the computer may be configured to extract ω1 and ω2 by calculating ωi=
−∥QCOD−γi∥β
where i={1, 2} and wherein γ and β are trained or empirically determined coefficients.
Moreover the invention relates to a computer program for speech quality estimation. The computer program comprises code means which when run on a computer connected to a communications network causes the computer to:
Q
COD+ωi·BW+ω2·PL; and
The computer program may comprise code means which when run on the computer causes the computer to extract ω1 and ω2 by calculating ω1 and ω2 according to
where i={1, 2} and γ, α and β are trained or empirically determined coefficients.
The computer program may comprise code means which when run on the computer causes the computer to determine QCOD by extracting QCOD from
wherein N is a number of frames or blocks in the speech signal and W is a number of frequency bands wherein the N and the W are related to a codec bit rate with n being a time frame, frame index or frame counter value and f being a frequency counter or band index value, and P represents power spectrum of the speech signal.
Furthermore the invention relates to a computer program product comprising computer readable code means and the computer program, which is stored on the computer readable means.
The objects, advantages and effects as well as features of the present invention will be more readily understood from the following detailed description of exemplary embodiments of the invention when read together with the accompanying drawings, in which:
a shows another embodiment of the speech quality estimation system.
While the invention covers various modifications and alternatives, embodiments of the invention are shown in the drawings and will hereinafter be described in detail. However it is to be understood that the specific description and drawings are not intended to limit the invention to the specific forms disclosed. On the contrary, it is intended that the scope of the claimed invention includes all modifications and alternatives thereof falling within the spirit and scope of the invention as expressed in the appended claims.
Presentation level variations and bandwidth limitations are typical distortions in a speech communication system/telecommunication network. In the presence of coding distortions, relation between the bandwidth and the presentation level degradations and perceived quality becomes non-linear. This is illustrated in
MOS is a listening test described in [8] ITU-T Rec. P.800 (August 1996), Methods for Subjective Determination of Transmission Quality. Listeners grade the signal quality on a scale 1 to 5, with the meaning 1 (bad), 2 (poor), 3 (fair), 4 (good), 5 (excellent). MNRU is a method to introduce controlled degradation in the speech signals, typically used as an anchor condition in listening tests. The speech signal is degraded by mixing it with a speech correlated noise, at a pre-defined level. Perceptually it mimics the effect of quantization noise, introduced by the speech compression system. The method is described in [9] ITU-T P.810 (February 1996), Telephone Transmission Quality, Methods for Objective and Subjective assessment of Quality, Modulated Noise Reference Unit (MNRU).
In the existing solutions mentioned above, the non-linear interactions between different quality dimensions is either not captured (documents [2]-[5]), or blindly modeled by means of artificial neural networks as in document [6]. Ignoring these effects or even using a simple linear model does not work, as illustrated in
It is therefore suggested according to the invention an inclusion of a bandwidth related distortion parameter (BW) and a presentation level distortion parameter (PL) in a speech quality estimation measurement. This inclusion preserves much of the linear model/modeling possibility, which in turn provides enhanced stability in speech quality estimation systems. The BW and the PL contribute to the general quality of a signal quality measure (Q) in a semi-linear model, with coefficients ωi where i={1, 2} dependent on the level of a coding distortion parameter QCOD, see Equation 1 and 2.
Q=Q
COD+ω1BW+ω2PL (1)
Here the coefficients γi, βi and αi are coefficients trained against subjective data/empirically determined e.g. by quality grades from listening test. The range for the coefficients ω1, ω2 depends on the range of QCOD, the PL and the BW. As an example, if {QCOD, PL, BW} are between 0 to 1; then the coefficients ω1, ω2 may be between −1 to 1. The coefficients ω1, ω2 are optimized to maximize prediction accuracy between an original quality and a predicted quality. The optimization can be performed in different ways known to the skilled person, but an example is to minimize the mean square error between objective quality and subjective quality, where the objective quality is a value retrieved from a computation by a computer and the subjective quality is a value retrieved via tests where humans judge the quality.
From equation (2) one can see that bandwidth and the presentation level degradations can contribute positively or negatively, based on the level of coding noise. The coding distortion QCOD can be determined from the codec bit-rate, perceptual model such as PESQ in document [2], or measured directly on the speech signal, e.g., through an average spectral flatness, see equation (3).
The QCOD might represent an overall coding distortion, or just a certain quality dimension, like noisiness, spectral outliers, etc. In Equation 3, N is a number of frames/blocks in the speech signal and W is a number of frequency bands wherein the N and the W are related to a codec bit rate with n being a time frame/frame index/frame counter value and f being a frequency counter/band index value, and P represents power spectrum of the speech signal.
The Q 530 value can be used to:
a shows another embodiment of the speech quality estimation system 500. In the telecommunications network 540, the OS 510 may be transcoded/altered at different sub-systems/network nodes i.e. N1, N2, . . . Nm and consequently the PS1, PS2, . . . PSm generated signals may be fed into the computer 700. This resulting in the Qj 530 (where j=1, 2, . . . m), i.e. for the different/individual sub-systems i.e. N1, N2, . . . Nm of the telecommunications network 540. So the OS 510 is fed into the SQES and also fed into the sub-system N1 of the telecommunications network 540. The output Q1530 then is measure of signal quality for the sub-system N1 of the telecommunications network 540. This can be repeated for the sub-systems N2 . . . Nm. The flow 600 below describes that the steps involved in the Q 530 generation may include the repeat procedure for the sub-systems described above in conjunction with
Although the respective unit disclosed in conjunction with
Furthermore the SQES comprises at least one computer program product 710 in the form of a non-volatile memory, e.g. an EEPROM (Electrically Erasable Programmable Read-only Memory, a flash memory and a disk drive. The computer program product 710 comprises a computer program 711, which comprises code means which when run on the SQES causes the SQES to perform the steps of the procedures described above in conjunction with
Although the code means in the embodiment disclosed above in conjunction with
The presented scheme for incorporating effects of the BW and the PL degradations allows keeping a semi-linear model in the quality assessment algorithm, which guarantees stable performance with unknown data. The presented scheme can be used as an extension to any of the existing standards for speech quality assessment such as the PESQ in document [2], PEAQ (Objective Measurements of Perceived Audio Quality) in document [6], MNB (Measuring Normalizing Block) in document [4] and P.563 in document [5].
A further embodiment of the invention is a method for a speech quality estimation system, comprising a speech quality estimation computer, e.g. in the form of a SQES. The method comprises steps, performed by the speech quality estimation computer, of:
Q
COD+ω1BW+ω2·PL
For a positive ω1, ω2 value, the Q of said signal improves/increases as the sum of distortion decreases. For a negative ω1, ω2 value, the Q of said signal decreases/degrades as the sum of distortion decreases.
In another embodiment of the invention, there exist provisions for an arrangement comprising a speech quality estimation computer, e.g. a SQES, adapted for being connected to a communications network. The speech quality estimation computer comprises:
Q
COD+ω1·BW+ω2·PL
In another embodiment of the invention, there exists provisions for a computer program for a speech quality estimation, the computer program comprises code means which when run on a speech quality estimation computer connected to a communications network, causes the speech quality estimation computer to:
Q
COD+ω1·BW+ω2·PL
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/SE2010/050867 | 7/26/2010 | WO | 00 | 1/19/2012 |
Number | Date | Country | |
---|---|---|---|
61228212 | Jul 2009 | US |