DE69033011T2 - DIGITAL VOICE DECODER USING A RE-FILTERING WITH A REDUCED SPECTRAL DISTORTION - Google Patents

Abstract

An adaptive spectral postfilter (108) in a synthesized speech platform has a denominator characteristic that corresponds to a preceding LPC filter stage (106) and a numerator characteristic that is developed as a function of the denominator characteristic through application of spectral smoothing techniques. This allows the numerator to track the denominator without the introduction of spectral distortion that would otherwise affect the processing in an adverse way.

Description

Technical area

This invention relates generally to speech coders and more particularly to digital speech coders that use post-filters to improve speech quality.

Background of the invention

Speech coders and decoders are well known in the art. Some speech coders convert analog voice samples into digitized representations and subsequently represent the spectral speech information by using linear predictive coding (see e.g. document EP-A-0 294 020). Other speech coders improve upon conventional linear predictive coding (LPC) techniques by providing an excitation signal related to the original voice signal.

US Patent No. 4,817,157 describes a digital speech coder and decoder having an improved vector excitation source, accessing a codebook of codebook excitation vectors to select a codebook excitation signal that is best suited to the available information and used to provide a synthesized speech signal from an LPC filter that closely represents the original.

Once the synthesized speech signal has been developed, various post-LPC filters are often used to further condition the signal. One of these filters is an adaptive spectral post-filter (which is typically used to improve the perceptual quality of the synthetic speech), and another is a post-emphasis filter (which adds brightness to the synthetic speech result).

An adaptive spectral postfilter typically has the general form:

where 0 ? ? ? &nu; < 1

and 1/1 - A(z) represents the associated LPC filter.

The denominator term in the above post-filter representation emphasizes the formants (partials) in the synthetic signal spectrum while attenuating the spectral valleys. (For the two extreme cases, setting ν = 0 results in an all-pass filter, while setting ν = 1 results in a denominator term that is the same as the associated LPC filter.) The numerator term attempts to constrain the general spectral shape introduced by the denominator. In state-of-the-art applications, ν is often set to approximately 0.8 and η to 0.5.

In practice, the numerator polynomial is only partially successful in tracking the spectral shape of the denominator (in fact, the spectral properties of the filter tilt over time), and the discrepancy typically reveals itself as a time-varying modulation of the postfiltered speech brightness.

Accordingly, there is a need for a method of post-filtering synthesized speech that both improves the perceptual quality of the synthetic speech and at the same time minimizes harmful effects on speech brightness. Preferably, speech brightness itself should be controlled better than before.

Summary of the invention

According to the present invention, there is provided a method for generating a synthesized speech signal, which comprises the following steps:

Providing an excitation signal to a linear predictive coding (LPC) filter; providing a synthesized speech signal from the LPC filter; providing a speech synthesis post-filter requiring a first component and a second component; providing the first component including a first set of coefficients; the method being characterized by the steps of: converting at least some of said first set of coefficients into an autocorrelation domain set of parameters; spectrally smoothing the autocorrelation domain set of parameters to provide a modified first set of coefficients; using the modified first set of coefficients to provide the second component for using the speech synthesis post-filter; filtering the synthesized speech signal in the speech synthesis post-filter using the first component and the second component to provide a filtered synthesized speech signal, the second component adaptively following and delimiting a general spectral shape of the first component; and acoustically reproducing the filtered synthesized speech signal.

In a first embodiment, Z-transform (filter) coefficients representing the first component are converted to the autocorrelation domain. A spectral smoothing technique using a bandwidth expansion function is then applied to the autocorrelation sequence, and the polynomial coefficients of the second component are calculated from the modified autocorrelation sequence via Levinson recursion. The first component is then used as the denominator and the second component as the numerator in the filter characteristic given above.

Through this process, the numerator polynomial is replaced by a spectrally smoothed version of the A(z/ν) polynomial. The bandwidth expansion of the formant does not alter the smoothed spectral envelope. Therefore, the spectrally smoothed bandwidth expanded version of the A(z/ν) polynomial effectively minimizes the time-varying spectral slope and allows the numerator to closely track and limit the general spectral shape of the denominator.

In another embodiment, an additional post-emphasis filter may be used to provide more control over post-filtered speech brightness. This filter is a first order filter of the form

(z) = 1 -uz⊃min;¹,

where typically 0.2 ≤ u ≤ 0.5.

An exemplary embodiment of the invention will now be described with reference to the accompanying drawings.

Short description of the drawings

Fig. 1 comprises a diagrammatic block representation of a radio device configured in accordance with the invention; and

Figure 2 is a flow chart illustrating the characterization of an adaptive spectral postfilter according to the present invention.

Best way to carry out the invention

U.S. Patent No. 4,817,157, entitled "Digital Speech Coder Having Improved Vector Excitation Source," issued March 28, 1989 to Ira Gerson, describes in particular detail a digital speech coder and decoder. As detailed in the above reference, this invention may be embodied in a speech coder (or decoder) using a suitable digital signal processor, such as a device in the Motorola DSP56000 family.

In Fig. 1, a radio device 100 embodying the invention is shown, which has an antenna 102 for receiving a voice-coded radio frequency (RF) signal 101. An RF unit 103 processes the received signal to recover the speech encoded information. This information is provided to a parameter decoder 105 which develops control parameters for various subsequent processes. An excitation source 104, as described above, uses the parameters provided to it to generate an excitation signal. The resulting excitation signal from the excitation source 104 is provided to an LPC filter 106 which produces a synthesized speech signal according to the encoded information. The synthesized speech signal is then pitch post-filtered 107 and spectrally post-filtered 108 to improve the quality of the reconstructed speech. If desired, a post-emphasis filter 109 may also be included to further enhance the resulting speech signal. (Additional details concerning the spectral post-filter 108 and the post-emphasis filter 109 are provided below.)

The speech signal is then processed in an audio processing unit 111 and reproduced audibly by a sound transducer. The excitation source 104, the LPC filter 106, the pitch post-filter 107, the matched spectral post-filter 108 and the post-emphasis filter 109 can all be provided by programming a DSP 113.

According to this invention, the adaptive spectral postfilter 108 is characterized by a first component (a denominator related to the filter characteristics of the LPC filter 106) and a second component (a numerator that follows the general spectral shape of the denominator to limit it). The general form of such a filter is described in an article entitled "Real-Time Vector APC Speech Coding at 4800 bps With Adaptive Postfiltering". Vector APC speech coding at 4800 bps with adaptive post-filtering) by Chen and Gersho, which appeared in the April 1987 issue of the Proceedings of the International Conference on Acoustics and Signal Processing, pages 2185-2188.

According to this invention, the numerator is developed by applying spectral smoothing techniques to the denominator polynomial. These techniques are described in an article entitled "Spectral Smoothing Technique in PARCOR Speech Analysis - Synthesis," by Tohkura, Itakura, and Hashimoto, which appeared in the December 1978 issue of I. E. E. E. Transactions on Acoustics, Speech, and Signal Processing.

In one embodiment, the Z-transform coefficients representing the denominator are converted to the autocorrelation domain. (Examples of such a conversion can be found in Markel, J. D.; Gray, A. H., Jr.; Linear Prediction of Speech; Springer-Verlag, Berlin, Heidelberg, New York, 1976.) The spectral smoothing and bandwidth expansion function is then applied to the autocorrelation sequence, with the numerator polynomial coefficients calculated via Levinson recursion from the modified autocorrelation sequence. In one embodiment, the autocorrelation coefficients are multiplied by the following factors to provide the resulting numerator coefficients:

Autocorrelation delay Spectral smoothing factor

0 1.0000000

1 0.9230769

2 0.7252747

3 0.4835164

4 0.2719780

5 0.1279896

6 4,9773753E-02

7 1,5718028E-02

8 3,9295070E-03

9 7,4847753E-04

10 1,0206513E-04

The denominator and numerator are then used to characterize the matched spectral postfilter 108.

It would of course also be possible to use the LPC filter information directly and develop the numerator term from it by a similar procedure, since the LPC filter information is used to develop the denominator term as described above.

Through this method, the numerator polynomial is provided by a spectrally smoothed version of the denominator polynomial. The spectrally smoothed bandwidth-extended version of the denominator polynomial effectively minimizes the time-varying spectral slope and allows the numerator to closely follow and limit the general spectral shape of the denominator. Based on listening tests, a bandwidth-extension factor (which specifies the amount of smoothing performed on the denominator) of approximately 1,200 Hz was used.

The flow chart in Fig. 2 helps to understand the postfilter characterization process now described. As previously explained, the matched spectral postfilter is characterized by a first component, or denominator, and a second component, or numerator. The first component is provided in block 202 and can be expressed by:

1 - A(z/ν).

In the subsequent step 203, the Z-transform coefficients are converted to the autocorrelation domain. In block 204, a spectral smoothing and bandwidth expansion function is applied to the autocorrelation sequence, and in the subsequent block 205, the numerator polynomial coefficients (second component) are calculated from the autocorrelation sequence modified in the previous step 204 by using Levinson recursion. The numerator, or second component, can be expressed by:

1 - B(z).

Finally, in step 206, the first and second components (denominator and numerator) are used to characterize the matched spectral postfilter, which can be represented as:

The post-emphasis filter 109 may be provided to provide more control over the post-filtered speech brightness. This filter is a first order filter of the form

(z) = 1 - uz⊃min;¹

Claims (6)

1. A method for generating a synthesized speech signal, comprising the following steps: A) Providing an excitation signal to a linear predictive coding filter; B) providing a synthesized speech signal from the linear predictive coding filter; C) providing a speech synthesis post-filter requiring a first component and a second component ; D) providing the first component containing a first set of coefficients; E) converting at least some coefficients of this first set of coefficients into an autocorrelation domain set of parameters; F) spectral smoothing of the autocorrelation domain set of parameters to provide a modified first set of coefficients; G) using the modified first set of coefficients to provide the second component for use by the speech synthesis post-filter; H) filtering the synthesized speech signal in the speech synthesis post-filter using the first component and the second component to provide a filtered synthesized speech signal and; I) audible reproduction of the filtered synthesized speech signal. 2. A method for generating a synthesized speech signal according to claim 1, further comprising the following steps: A) Receiving a radio frequency (RF) signal containing encoded voice information; B) recovery of an excitation signal from the coded speech information; C) providing the excitation signal to a linear predictive coding (LPC) filter; and wherein a first component is provided for use by the speech synthesis post-filter. 3. The method of claim 1 or 2, wherein the LPC filter is at least partially defined by the expression: 1/1 - A(z) 4. The method of claim 1 or 2, wherein the first component of the speech synthesis post-filter is of the form: 1 - A(z/&nu;) when it is represented in Z-transform notation . 5. The method of claim 1 or 2, further comprising the following steps: A) Filtering the synthesized speech signal in a post-emphasis filter, which in Z-transform notation is essentially defined by: (z) = 1 - uz⊃min;¹ where 0.2 ≤ u ≤ 0.5. 6. A method according to any preceding claim, wherein the step of calculating includes the step of multiplying.