DE4236315C1 - Method of speech coding - Google Patents

Abstract

The invention concerns a speech-encoding method using analysis-by-synthesis techniques. The speech signal is scanned, a frame formed from a predetermined number of scanning samples, and the coefficients of a grade P speech-synthesis filter determined from the samples in each frame. Using these coefficients, a number P of so-called line-spectrum parameters (LSPs) are determined and quantized, for transmission over a channel with limited transmission capacity. The method is characterized in that every other line-spectrum parameter (LSP) is quantized in scalar (absolute) fashion and that the line spectrum parameters (LSPs) lying between them are transformed (normalized) before quantization. The invention is intended particularly for use in speech-encoding equipment in portable radio equipment.

Description

The invention is based on a method for speech coding using the analysis-by-synthesis method according to The preamble of claim 1. Such speech coding method are known, for example by the German patent 38 34 871.

Common to the speech coding method is a Prediction analysis of the input signal (Linear Prediction Coder, LPC). The voice signal at the input of the Encoders within a certain duration of z. Eg 20-30 ms divided. Each speech frame is linear in the encoder Subjected to prediction analysis, which linear dependencies removed in the voice signal. The linear prediction is with Help with FIR (Finite Impulse Response) filters. The coefficients of these linear filters are in each Frame redetermined, d. h., this is adaptive Filter.

Today 's speech coders, which operate at bit rates between 4 and 16 kBit / sec, usually use the analysis by-synthesis method, wherein in the transmitter the above Filter coefficients and a corresponding excitation so determined that the energy of the weighted error e (n) between the original language and the synthesized language as small as possible.

Must be transmitted to the receiver parameters that the Describe the suggestion, and those already mentioned above Coefficients of the linear filter. On the determination of Coefficients of the linear filter should not be closer To be received. The result is one  nonrecursive filter of grade P with the transfer function

The inverse transfer function H (z) -1 / A (z) converts this Error signal (the excitation) into the (synthesized) Speech signal around:

The filter H (z) calculated by this method is stable in each case without quantization of the filter coefficients a _i .

However, the filter coefficients a _i have a large dynamic and are therefore poorly suited for quantization and transmission. In addition, there is no easy way to check the stability of the recursive filter in the receiver.

It is known that the so-called line-spectrum parameters LSP for quantization and transmission, ie for Description of the predicator H (z) are suitable. you gets these parameters as zeros of a symmetric one polynomial

F₁ (z) = A (z) + Z ^{- (P + 1)} A (z ^-1 )

and an antisymmetric polynomial

F₂ (z) = A (z) - Z ^{- (P + 1)} A (z ^-1 ).

The zeros z _0i of F₁ and F₂ have the following properties

• - all zeros are on the unit circle, so are described by the specification of a phase ω _i sufficient,
• - all zeros are simple,
• - On the unit circle is alternately a zero of F₁ and F₂.

In Fig. 2, the zeros of F₁ (z) and F₂ (z) for the cases P = 6 and P = 5 are shown. All zeros z _i can be determined by the arguments ω _i or by the derived frequency value

being represented.

Since the zeroes occur in conjugate complex pairs and zeroes at ± 1 are present in each case, the polynomials F₁ and F₂ are completely determined by the specification of P values ω _i .

After the properties described above must apply

ω₁ <ω₂ <. , , <ω _P.

This monotony property is imperative for that recursive filter H (z) is stable. It has a criterion to check the stability of the filter.

When changing the characteristic of the spectrum of the input signal, the distribution of individual LSPs changes greatly. As an example, the distribution of the LSP for filter degree P = 10 is shown in FIG . In the upper picture, Fig. 1a, the input speech is only low-pass filtered, in the lower picture, Fig. 1b, IRS filtered (band-limited) to CCITT P.48.

A common technique is the scalar quantization of each individual LSPs, for example, in 4.8 kbit / sec CELP Language code according to the Federal Standard 1016 of the US US Department of Defense the Line Spectrum parameters  scalar quantized with a total of 34 bits.

It should be noted in the quantization that even after the Quantization the monotonicity property must be obtained so that the recursive filter is stable; d. h., it must apply:

Since the value ranges of the quantizers overlap for ω _i and 2 _{i + 1} , after the quantization of ω _i, all quantization stages of ω _{i + 1} which violate this severe monotony are excluded (see FIG. 3). Conversely, even after the quantization of ω _{i + 1} , values from the quantizer ω _{i are} no longer permissible. This means that some of the bits available for quantization of the LSP parameters are not fully utilized. Of Fig. 3 of 8 possible stages for ω _{i + 1} actually only 5 are used.

Another disadvantage of this method is that an adaptation not to different input spectra of the speech signal is possible. Should the quantizer be used for this purpose can increase the value range of individual line Spectrum parameter. This leads to an increase in the bit rate.

In references [5] and [6], a reduction of the Bitrate for transmission of line spectrum parameters Quantization whose differences are proposed. It is the first LSP scalar scalarized as above

For all other LSPs the difference to the previous one becomes Value calculated and then quantized.

This method adapts well to different input spectra of the speech signal, since only the value range of the first LSP must be chosen to be sufficiently large. A disadvantage of this method is the propagation of errors. If an error occurs during the transmission of _x , all _i for i = x to P are decoded incorrectly.

By the European patent application EP 0 483 882 A2 is a Speech parameter coding method for coding a Input speech signal into a sequence of coded signals for transmission of spectrum parameters with a reduced Number of bits have become known. This method uses a multilevel vectorial quantization of the line-spectrum Parameter. Vector quantization becomes a reduction the bit rate achieved, but at the expense of a Significant increase in memory and computational effort goes.

The present invention has the object, a Specify method of the type mentioned, which in the Location is at constant bitrate an improvement in the To achieve voice quality or at the same time Voice quality to achieve a reduction in bit rate. In addition, a reduction in the sensitivity of the Language codes versus speech signals with different Input characteristics can be achieved. The needed Circuit complexity should not be too high.  

This object has been achieved by the features of the claim 1. Advantageous embodiments result from the Dependent claims.

The inventive method achieves the advantages of Improvement of the voice quality with constant bit rate or a reduction in the bit rate at the same Voice quality. In addition, the inventive method a reduced sensitivity of the language code Speech signals with very different input spectra on. Another advantage is that a Transmission error in a LSP only to a maximum of two more Affects LSP values.

The invention is based on the idea of not all LSP parameters scalar quantize only a single of the total P Quantize parameters scalarly, but only every nth of the P parameter scalar quantize and the intervening To transform or map parameters and then quantize.

The method is described below with reference to a Embodiment described in more detail, where adopted is that P is an even number.

In a first step every second LSP becomes scalar quantized.

Now it must be due to the strict monotony

wherein the notional value ω _{P + 1 is set} to the maximum possible value for ω _P. This value range for ω _i changes from frame to frame with and. It would be ideal to use for each combination of and a separate quantizer for ω _i , which is not possible for reasons of implementation effort. Instead, the value range is mapped to the interval [0, 1] by the following transformation:

Each value x _i can now be quantized and transmitted with a quantizer. The inverse transformation takes place according to

The procedure works accordingly, if one the Parameters that are absolutely quantized with those which are quantized after normalization, d. H.

absolutely quantize: ω _i i = 2, (2), P quantize after transformation: ω _i i = 1, (2), P - 1

Instead of transforming the LSP into the image area it is also possible to use the quantizer from the image area according to (13) in the ω-range.

Similarly, in the second embodiment every third LSP scalar quantized.

The mapping function for the intermediate parameters are for example

or

since ω _{i is} now known.

This solution provides a further reduction of Bitrate with simultaneous quality or higher Quality at constant bit rate; however, it does affect a transmission error here to max. three more LSP values out.

In a similar way can also be proceeded by only every fourth LSP scalar quantized and the intervening ones LSP transformed accordingly and then quantized be transmitted.

Claims (3)

A method of speech coding using the analysis by synthesis method, wherein the speech signal is sampled, a frame is formed from a fixed number of samples and the coefficients of a speech synthesis filter of degree P are determined frame by frame from the samples By means of these coefficients, a number P of so-called line-spectrum parameters LSP are determined and quantized, for transmission over a channel with limited transmission capacity, characterized in that every n-th line-spectrum parameter LSP is scalar (absolute) quantized and that the intermediate line spectrum parameters LSPω _i for i = 1, P and im, (n), P transformed (normalized) and then quantized. 2. The method of claim 1 with n = 2, characterized in that the transformation according to the function he follows. 3. The method of claim 1 with n = 3, characterized in that the transformation according to the mapping functions or he follows.