EP0484339B1

EP0484339B1 - Digital speech coder with vector excitation source having improved speech quality

Info

Publication number: EP0484339B1
Application number: EP90908908A
Authority: EP
Inventors: Ira Alan Gerson
Original assignee: Motorola Inc
Current assignee: Motorola Solutions Inc
Priority date: 1989-06-23
Filing date: 1990-05-02
Publication date: 1998-02-04
Anticipated expiration: 2010-05-02
Also published as: WO1991001545A1; KR950003557B1; IL94119A0; IL94119A; KR920702787A; EP0484339A4; AU638462B2; AU5735990A; BR9007467A; EP0484339A1; DE69032026D1; DE69032026T2; NZ234180A; CA2060310A1; CA2060310C; CN1023160C; CN1048278A

Abstract

In a vector excitation source digital speech coder utilizing vector excitation, candidate excitation sources (111, 121) are considered independent of certain pitch parameters. Once a particular excitation source has been selected, the excluded pitch parameter may then be optimized, resulting in an overall improvement in speech quality.

Description

Technical Field

This invention relates generally to speech coders, and more particularly to digital speech coders that use vector excitation sources.

Background of the Invention

Speech coders are known in the art. Some speech coders convert analog voice samples into digitized representations, and subsequently represent the spectral speech information through use of linear predictive coding. Other speech coders improve upon ordinary linear predictive coding techniques by providing an excitation signal that is related to the original voice signal. Previously issued U.S. Patent No. 4,817,157 describes a digital speech coder having an improved vector excitation source wherein a codebook of excitation vectors is accessed to select an excitation signal that best fits the available information, and hence provides a recovered speech signal that closely represents the original.

In general, the resultant decoded speech signal will more closely represent the original unencoded speech signal if there is a significant number of candidate excitation vectors available for consideration as the excitation source. Increasing performance in this way, however, generally results in enlargement of the codebook size, and this will usually increase processing complexity and data rates.

P. Kabal et al., in "Synthesis Filter Optimization and Coding: Applications to CELP", ICASSP88 (1988 International Conference on Acoustics, Speech, and Signal Processing, New York, 11-14 April 1988) Vol. 1, p. 147-150, IEEE, New York, USA), describe a coder with a reduced codebook size (16 or 32 waveforms) to overcome this problem. For best results, Kabal et al recommend joint optimisation of the pitch lag and the candidate vector instead of first determining the pitch lag and then (sequential approach) identifying the most appropriate candidate vector. This joint optimisation is computationally complex but desirable for accurate coding because the candidate vectors used by Kabal et al may include information which is dependent on the pitch lag. According to Kabal et al, the sequential approach is computationally simpler, but the quality of the decoded speech signal is not as good.

A need therefore exists for a digital speech coder that uses a vector excitation signal, wherein for a given size codebook, the quality of the decoded speech signal is substantially maximized with minimal increase in complexity and substantially no increase in data rate.

Summary of the Invention

These needs and others are substantially met through provision of the digital speech coder with vector excitation source having improved speech quality disclosed herein.

Accordingly, there is provided a method of encoding a speech sample as set forth in claim 1 and a method of encoding a signal sample as set forth in Claim 7.

This methodology allows candidate excitation signals to be considered without requiring a commensurate increase in processing complexity or data rates.

The coded excitation signal is determined substantially independent from any pitch information. In particular, candidate excitation signals as provided by a codebook are processed to substantially remove components that are representable, at least in part, by a reference component that is related, at least in part, to the intermediate pitch vector. More particularly, the vector component related to the intermediate pitch vector is removed from the candidate excitation signal (a process known as orthogonalizing). The orthogonalized candidate excitation signals are then compared with the unencoded speech sample to identify the candidate excitation signal that best represents this particular speech sample. The pitch information, including a pitch filter coefficient parameter, can be optimized later to best suit the selected excitation signal to thereby yield an overall optimized coded representation of the speech signal.

In another embodiment, a second codebook of candidate excitation signals, wherein two excitation signals are used to represent the speech sample, is provided. The first excitation signal can be selected as described above, and the second excitation signal can be selected in a similar manner, wherein candidate second excitation signals are first orthogonalized with respect to both the intermediate pitch vector and the previously selected first excitation signal.

Brief Description of the Drawings

Fig. 1 comprises a block diagrammatic depiction of the invention; and

Fig. 2 comprises a simple vector diagram representing one aspect of the invention.

Best Mode For Carrying Out The Invention:

This invention can be embodied in a speech coder that makes use of an appropriate digital signal processor such as a Motorola DSP 56000 family device. The computational functions of such a DSP embodiment are represented in Fig. 1 as a block diagram equivalent circuit.

A pitch period parameter (101) (determined in accordance with prior art technique) is provided to a pitch filter state (102) that comprises part of a pitch filter. The resultant signal (103) comprises an intermediate pitch vector that is provided to both a first multiplier (104) and two orthogonalizing processes (106 and 107) as described below in more detail. This first multiplier (104) functions to multiply the resultant signal by a pitch filter coefficient (108) to yield a pitch filter output (109). Selection of the pitch filter coefficient (108) will be described below in more detail.

A first codebook (111) includes a set of basis vectors that can be linearly combined to form a plurality of resultant excitation signals. Depending upon the size of the memory utilized, and other factors appropriate to the application, the number of possible resultant excitation signals can be, for example, between 64 and 2,048, with more of course being possible when appropriate to a particular application. The problem, when encoding a particular speech sample, is to select whichever of these excitation sources best represents the corresponding component of the original speech information.

Pursuant to this invention, once a particular resultant signal (103) has been determined, the excitation signals formulated by the first codebook (111) will be presented in seriatim fashion as candidate excitation sources. Each candidate excitation source will first be orthogonalized (106) with respect to the resultant signal. For example, referring momentarily to Fig. 2, if vector A were considered to represent the resultant signal and vector B were to represent a particular candidate excitation source, orthogonalization of the candidate excitation source signal would result in the vector denoted by reference character B'. (It should be understood that in practice, the vector dimension space is a function of the number of samples comprising the vectors, which may be upwards of 40 samples or more. It should also be noted that the candidate excitation vectors may be readily orthogonalized by orthogonalizing the basis vectors, wherein linear combinations of the orthogonadized basis vectors with one another will result in orthogonalized excitation vectors.)

Once orthogonalized, the resulting candidate excitation source can be compared (112) with the unencoded signal (113) (or an appropriate representative signal based thereon) to determine the relative similarity or disparity between the two. The process is then repeated for each of the excitation sources of the first codebook (111). A determination can then be made as to which candidate excitation source most closely aligns with the unencoded signal (113).

In this particular embodiment, a gain factor (114) can also be used to modify each candidate excitation source signal, as well understood in the art. In addition, if desired, the excitation source selection and gain compensation can both be accomplished in a substantially simultaneous manner, as also well understood in the art.

Once an appropriate excitation source from the first codebook (111) has been selected through this process, the orthogonalizing process (106) can thereafter be dispensed with and the exact excitation source signal selected (116) through an appropriate control mechanism (117). Thereafter, presuming a single codebook coder, the pitch information can be gated (117) and summed (118) together with the selected excitation source with the pitch filter coefficient (108) and excitation gain (114) optimized such that the combined excitation most closely aligns with the encoded signal (113). Once optimized, the pitch period parameter, pitch filter coefficient, and particular excitation source and gain are known, and appropriate representations thereof may be utilized thereafter as representative of the original speech sample.

If desired, and as depicted in Fig. 1, an additional codebook (121) can be utilized, which second codebook (121) again includes a plurality of basis vector derived candidate excitation sources. The use of such multiple codebooks is understood in the art. Pursuant to this invention, however, once the first excitation source from the first codebook (111) has been selected as described above, the candidate excitation sources from the second codebook (121) are orthogonalized (107) with respect to both the resultant signal (103) and the selected excitation source signal from the first codebook (111). The selection process can then continue as described above, with the orthogonalized candidate excitation source signals from the second codebook (121) being compared against a representative unencoded signal (113) to identify the closest fit. Once this excitation source has been selected, the pitch filter coefficient (108) and excitation gains (114 and 120) can then be optimized as described above.

Claims

A method of encoding a speech sample, comprising the steps of:

A) determining a pitch period parameter for the speech sample;

B) determining, independent of any pitch filter coefficient, a plurality of coded excitation signals for the speech sample;

C) processing each of the plurality of coded excitation signals to provide a plurality of processed candidate excitation signals, wherein each of the plurality of processed candidate excitation signals are comprised of information that is substantially independent of information that is representable by a pitch filter output that is derived, at least in part, as a function of the pitch period parameter;

D) optimizing at least a pitch filter coefficient parameter for the speech sample.
The method of claim 1 further characterized in that the step of determining a coded excitation signal includes processing the plurality of candidate excitation signals to orthogonalize the plurality of candidate excitation signals with respect to a pitch filter output that is derived, at least in part, as a function of the pitch period parameter.
The method of claim 1, further characterized in that the step of determining the coded excitation signal comprises the steps of:

B1) processing an excitation signal to substantially remove components that are representable, at least in part, by a reference that is related, at least in part, to the pitch period parameter; and

B2) determining an appropriate excitation signal for the speech sample.
A method of claim 3 further characterized in that the step of processing the excitation signal includes processing the excitation signal to orthogonalize the excitation signal with respect to a pitch filter output that is derived, at least in part, as a function of the pitch period parameter.
The method of claim 3, and further characterized by the step of:
C1) processing a candidate excitation signal to substantially remove components that are representable at least in part, by:
a reference that is related, at least in part, to the pitch period parameter; and the appropriate excitation signal determined in step C.
The method of claim 5, further characterized in that the step of processing a candidate excitation signal includes processing the candidate excitation signal to orthogonalize the candidate excitation signal with respect to both the reference and the appropriate excitation signal determined in step C.
A method of encoding a signal sample using at least two codebooks that include information regarding candidate excitation signals, comprising the steps of:

A) determining, using a first one of the codebooks, a first excitation signal for the signal sample; characterized by:

B) determining, using a second one of the codebooks, a second excitation signal for the signal sample, which second excitation signal is comprised of information that is substantially independent of information that is representable by the first excitation signal;

C) using the first and second excitation signals to represent, at least in part, the signal sample.
The method of claim 7 further characterized in that the signal sample comprises a speech sample.
The method of claim 7 further characterized in that the step of determining the second excitation signal includes processing a candidate excitation signal to orthogonalize the candidate excitation signal with respect to the first excitation signal.