The present invention relates to a method for extending line spectral frequencies of a narrowband speech signal with a frequency range to line spectral frequencies of a wideband speech signal comprising a highband frequency range and the frequency range of the narrowband speech signal and to a system for extending the frequency range of speech signals at an input comprising an output and an upsampler connected to the input of the system and an input analysis means for determining linear prediction coefficients and reflection coefficients, an input of the input analysis means connected to the input of the system, the upsampler comprising an output connected to an input of a first filter, which first filter comprises an output and is arranged to filter based on linear prediction coefficients, the output of the first filter connected to a an input of a spectral folding means with an output connected to an input of a second filter comprising an output, which second filter is arranged to filter based on the linear prediction coefficients, the output of the second filter being connected to the output of the system for extending the frequency range of speech signals
Such a method and system is known from the publication ‘wideband extension of telephone speech using a hidden Markov model’ by Peter Jax and Peter Vary, IEEE Workshop on Speech coding, September 2000, Wisconsin. Here the narrowband input signal is classified into a limited number of speech sounds in which the information about the wideband spectral envelope is taken from a pre-trained code book. For the codebook search algorithm a statistical approach based on a hidden Markov model is used, which takes different features of the bandwidth limited speech into account, and minimizes a mean squared error criterion. The algortihm needs only one single wideband codebook and inherently guarantees the tranparency of the system in the narrowband frequency range. The enhanced speech exhibits a significant larger bandwidth than the input speech. The algortihm creates the entire wideband signal by applying codebook LPC coefficients to a first, inverse, filter that acts on the input signal and then provides the filtered and subsequently spectrally folded signal to a second, synthesis, filter. This synthesis filter also receives codebook LPC coefficients and provides the wideband signal at the output. Because the transfer functions of these two filters are mutually inverse the narrowband signal is processed transparently by the system.
This method of wideband extension has the disadvantage that the filtered signal as provided by the first filter is not sufficiently flat to provide, after spectral folding, an optimal signal for the second filter to create a highband speech signal.
The objective of the present invention is to provide a method of extending a narrowband speech signal to a wideband speech signal where after spectral folding an optimal signal is provided to the inverse filter.
The invention achieves this object by applying the following steps
Deriving line spectral frequencies for the extended frequency range of the wideband speech signal by applying a matrix obtained by training to line spectral frequencies of wideband speech signals in the frequency range of the narrowband speech signal.
Mapping the line spectral frequencies of the narrowband speech signal to line spectral frequencies of the wideband speech signal in the frequency range of the narrowband speech signal
Combining the line spectral frequencies for the highband frequency range with the line spectral frequencies of the narrowband speech signal.
This way the LSFs of the narrowband speech signal are mapped directly without processing to the equivalent lowband LSFs of the wideband speech signal, while the highband frequency range of the wideband signal is created by applying a matrix to the LSFs of the narrowband speech signal. Because the mapping of the highband LSFs does not affect the lowband LSFs, an optimally flat signal can be obtained from the first filter. After spectral folding, the spectrum of the folded signal remains flat providing an optimal input signal for the synthesis filter.
One method to obtain the highband LSFs is by applying a matrix obtained by training to line spectral frequencies of wideband speech signals in the frequency range of the narrowband speech signal. Also the use of multiple matrices to further optimize the synthesis of the highband signal is enabled by the independent processing.
The line spectral frequencies are obtained by decomposition of the impulse response of the LPC analysis filter into even and odd functions. In this extension technique LSFs are estimated from the input narrowband signal. The LSFs are located between 0-π in 4 kHz bandwidth of a narrowband speech signal sampled at 8 kHz. Assuming that the corresponding wideband speech is modelled using an LPC model with twice the order of the narrowband LPC model, the narrowband LSFs should represent the wideband LSFs in the lowband range 0-π/2. Thus the lowband LSFs of the wideband speech signal are given as the narrowband LSFs divided by 2.
In a simulation of the wideband speech where the synthesis uses lowband LSFs obtained from narrowband speech as described above and the highband LSFs are taken from the corresponding wideband speech very good output quality was obtained.
The high band LSFs can obtained from the lowband LSFs using a matrix. The matrix is obtained by training and needs to be established just once. It is also possible to obtain several matrices, each matrix being specific to the type of signal being processed. Once such a matrix is obtained the wideband LPC coefficients are obtained as follows:
First linear prediction and reflection coefficients of the narrowband speech signal are estimated. Then LSFs are computed from these linear prediction. These LSFs are divided by two and provided directly to an array appender and to the highband LSF estimator. The highband LSF estimator applies a matrix selected from a set of matrices to the divided LSFs. The matrix selection is based on the type of signal that is being processed.
The result of the application of the selected matrix to the divided LSFs is a set of highband LSFs. These highband LSFs are then provided to the array appender. The array appender appends the highband LSFs to the lowband LSFs to form the wideband LSFs. The resulting array of wideband LSFs allows the calculation of the wideband LPCs which are used in the synthesis of the wideband speech signal in a system such as disclosed by Jax. LSFs and LPC coefficients form the basis of various methods and systems for extending the frequency range of a speech signal that improve the perceived quality of said speech system. There fore the extension of narrowband LSFs and LPC coefficients to wideband LSFs and LPC coefficients as provided by the present invention can be used in other systems for extending the frequency range of a speech signal as well.
The extension of the frequency range of speech signals is used in receiving terminals in systems where channel resources are to be conserved and speech is transmitted with a narrow bandwidth. Examples of the systems include mobile phones, video conferencing terminals and internet telephony terminals.
The present invention will now be described based on figures.
The first two reflection coefficients k1, k2 of all the reflection coefficients provided by the input analysis means 3 are used to classify the speech signal by determining to which cluster of reflection coefficients the reflection coefficients k1 and k2 are associated. Based on a search, for instance a bayesian search, by the matrix selector 15 a matrix M is selected from a matrix list 17 of predetermined matrices. These predetermined matrices are obtained by training to line spectral frequencies of wideband speech signals in the frequency range of the narrowband speech signal.
The matrix selector 15 provides either the selected matrix or information indicating which matrix was selected to the highband LSF estimator 9 in
The matrix selector 59 provides either the selected matrix or information indicating which matrix was selected to the highband LSF estimator 9 in
The system for extending the frequency range of a speech signal of
The upsampler 71 provides an upsampled signal to the first filter 81. The first filter 81 then filters this upsampled signal where the filter uses the wideband LPC parameters as provided by the linear prediction determinator 13. The wideband LPC parameters are obtained in the same fashion as described in
The first, inverse, filter provides a filtered signal to the spectral folding means 85 where the frequency range of the filtered signal is extended by spectral folding. Since the filtered and spectrally folded signal is used by the synthesis filter 87 to create the wideband output signal using the wideband LPC coefficients it is important that the filtered signal at the output of the inverse filter is spectrally flat in order to ensure that after spectral folding the highband portion of the filtered signal remains spectrally flat before being filtered by the synthesis filter 87. By providing the lowband LSFs, after multiplying by 0.5, directly to the inverse filter 81 an optimal signal can be provided to the synthesis filter 87, resulting in an optimal highband signal in the wideband signal. The synthesis filter 87 filters the filtered and spectrally folded signal using the same LPC coefficients as the first filter and provides an output signal with an extended frequency range at the output of the system.
Number | Date | Country | Kind |
---|---|---|---|
002039378 | Nov 2000 | EP | regional |
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/EP01/13137 | 11/9/2001 | WO | 00 | 7/3/2002 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO02/39430 | 5/16/2002 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
6704711 | Gustafsson et al. | Mar 2004 | B2 |
Number | Date | Country | |
---|---|---|---|
20020193988 A1 | Dec 2002 | US |