WO1997049189A1

WO1997049189A1 - Fourth order digital noise shaper circuit

Info

Publication number: WO1997049189A1
Application number: PCT/US1997/010735
Authority: WO
Inventors: Alfredo Linz
Original assignee: Advanced Micro Devices Inc
Current assignee: Advanced Micro Devices Inc
Priority date: 1996-06-20
Filing date: 1997-06-20
Publication date: 1997-12-24
Anticipated expiration: 1998-12-20

Abstract

A digital noise shaper circuit is provided which includes a signal transfer function and a noise transfer function, with associated coefficients and gain scaling factors, which ensures a predetermined frequency response, signal passband and noise stopband, and which converts a multi-bit digital signal to a 1-bit digital signal.

Description

TITLE: FOURTH ORDER DIGITAL NOISE SHAPER CIRCUIT

Specification Background of the Invention

1. Field of the Invention.

This invention relates to a digital noise shaping circuit. More particularly, this invention relates to a fourth order digital noise shaping circuit for use in a digital-to-analog conversion circuit.

2. Brief Description of the Related Technology.

Typically, digital-to-analog converter circuits (DACs) operate by a resistor/capacitor divider method, or use a sigma-delta conversion method, to convert digital signals to analog signals. Sigma-delta DACs are preferred because of their inherent feasibility to be manufactured in integrated circuits. These DACs typically create 1-bit signals from multi-bit digital input signals. This quantization, if performed by merely truncating or rounding the multi- bit signal, would introduce a large amount of noise in the signal passband, thereby destroying the signal quality. Noise shaping pushes the quantization noise out of the signal passband. Also, the 1-bit signal can be converted to an analog signal with no linearity errors.

Summary of the Invention The present invention is for a fourth order digital noise shaping circuit which takes an oversampled multi-bit input signal from the output of an interpolator circuit and converts the multi-bit signal to a 1-bit signal while shaping the quantization noise according to a high-pass function. A signal transfer function (STF) and a noise transfer function (NTF) define the operation of the noise shaper circuit. The NTF is chosen so zeros are introduced inside the noise stopband, which is substantially equal to the signal passband, while also achieving a flat high-frequency response. Once the NTF is determined, the STF is determined. A set of feedback coefficients are used to multiply the 1-bit output signal and feed it back to earlier stages in the noise shaper circuit to achieve the proper frequency response. Another set of coefficients is used to produce local feedback around certain stages of the noise shaper. The coefficients are chosen to achieve equiripple quantization noise density in the passband and a flat stopband. Suitable scaling factors between stages are also used to make the circuit stable for a predetermined range of input amplitudes.

To achieve optimum efficiency, a hardware block within the digital noise shaper circuit includes two integrators and associated adders which are identically scaled. Thus, for a fourth order noise-shaper circuit, only two unique scaling factor/integrator combinations are required to perform the fourth order function, instead of four such combinations.

Brief Description of the Drawings Fig. 1 is a block diagram of a digital to analog converter block utilizing the noise shaper of the present invention;

Fig. 2 is a schematic representation of the present noise shaper; Fig. 3 is a signal flow graph (SFG) of the present noise shaper;

Fig. 4 is a plot of the poles and zeros in the s plane for the present noise shaper; Fig. 5 is a plot of the noise transfer function magnitude of the present noise shaper;

Fig. 6 is a plot of the poles and zeros in the z plane of the present noise shaper; Fig. 7 is a graph of the noise transfer function, in the 0 - 10 KHz range, of the present noise shaper;

Fig. 8 is a plot of the ideal and realizable zeros of the present noise shaper;

Fig. 9 is a plot comparing two embodiments of noise transfer functions; Fig. 10 is a plot of the noise and signal transfer functions of the noise shaper;

Fig. 11 is a plot of the signal transfer function magnitude and phase in the passband of the noise shaper; and

Fig. 12 is a graph of the group delay (sec.) of the noise shaper.

Detailed Description of the Preferred Embodiment

Figure 1 illustrates a block diagram of the playback path of an audio CODEC. A suitable audio CODEC is described in application Serial No. 08/333,467, entitled "Stereo Audio CODEC", filed November 2, 1994, assigned to the common assignee of the present invention and incorporated herein for all purposes. Noise shaper 802 may be a stand-alone circuit, or may otherwise be included in an audio processing circuit, such as the audio CODEC illustrated.

In operation, noise shaper block 802 of Fig. 2 takes at least one multi- bit audio signal, illustrated as in.l and in.2, quantizes each respective input signal, converting the respective input signal to a 1-bit output signal, while shaping the quantization noise associated with the respective signal according to a high-pass function. The block diagram implementation for the shaper 802, which is preferably a fourth order shaper, is shown in Fig. 2. The remaining description of noise shaper block 802 is directed to only one multi- bit digital input signal, although the processing of another is identical.

The 1-bit output signal 842 is input to integrators 822a and 822b. Integrators 822a and b must have suitable scaling factors on the input, to make the loop stable for a predetermined range of input amplitudes, as determined by the remainder of the digital path shown in Fig. 2. The simple additive noise model shown in Fig. 2 is used to represent the quantizer.

Two transfer functions are defined for the fourth order noise shaper circuit: a signal Transfer Function (STF) Y/X, where X is the digital audio input signal 842, and a noise Transfer Function (NTF) Y/E, where E is the quantization noise (modeled as additive, white, uniformly distributed noise). Once the NTF is fixed, the STF is also determined. Since the system is not a FIR filter, the response is no longer strictly phase-linear. The phase variation in the passband, however, is very small, on the order of about 0.011 degrees.

A signal flow graph (SFG) for noise shaper block 802 is shown in Fig. 3. The transfer functions are developed as follows: Forward Path Gains: The cumulative gains of all possible direct paths from input to output:

For X:

Loop Gains:

The gains of all closed loops.

Non-touching Loops:

The products of the gains of sets of loops without any common nodes are calculated. First, pairs of non-touching loops have to be identified. Then, triplets are found, then sets of 4, etc. In the preferred embodiment, only pairs of non-touching loops exist.

Determinant: This is defined in terms of the loop gains as

In the preferred embodiment, there are no triplets of non-touching loops, so

Sub-determinants :

Δ_k = Δ setting to zero gains of loops touching forward path k For X:

AU loops are touched by Tl, so

Δ₁ = 1 For E:

Δ₁ = 1 — L₁ — L₂ + L₁L₂ The transfer functions can then be constructed for X and E using Mason's rule, where

The transfer functions have the form:

for noise, and

for the signal, where

Where, referring to Fig. 2,

The coefficients are chosen to match a Chebyshev function, which yields equiripple quantization noise in the passband and a flat stopband. The values for Ai and the Bi are obtained from the Ci and Wi in the above equations by matching the NTF to the desired shaping function.

Preferably, a function is chosen for the NTF which has zeros equally spaced inside the noise stopband (i.e., the signal band), and a flat high- frequency response. For the preferred embodiment, the stopband edge, the stopband attenuation and the filter order must be determined. Since the stopband attenuation is preferably at least 88 dB and the stopband edge is about 4 KHz for an input sampling rate of 8 KHz, or equivalently, about 24 KHz at the maximum sampling rate of 48 KHz, the filter order preferred is four.

First, the continuous time zeros and poles are obtained, where the zeros are given by:

and the poles by:

where N = 4, m ranges from 0 to 3, ω_r = stopband edge = 2π . 24000 at f. = 48 KHz, or 4 KHz at f. = 8 KHz, and el is related to the attenuation G given in dB by:

The pole-zero diagram in the s-plane is shown in Fig. 4. A plot of the frequency response out to 50 KHz is shown in Fig. 5, for f, = 8 KHz. Next, the discrete zeros and poles are obtained using the bilinear transformation:

k = 0,

where T = 1/f ., and f . = 64 x f. = 512 KHz, for f. = 8 KHz, where f. is the samphng rate of audio signal 806 and f , = 64 x f, is the sampling rate of interpolated signal 840. This is the highest sampling rate at which the noise shaper 802 will operate, and corresponds to an oversampling factor of 64 times the highest sampling rate for the input signal. It should be understood, however, that the noise shaper will be operated at other (lower) sampling rates.

Solving these equations yields:

Fig. 6 gives the pole-zero diagram in the z-plane for noise shaper 802.

K is the gain of the NTF at f = ϊJ2 (or z = -1) and is an important parameter for stabihty. The preferred frequency response of the discrete filter for noise shaper 802 is shown in Fig. 7.

The numerator in the transfer function of the selected structure must be matched to the discrete filter. The nature of the zeros that can be realized with it are found by equating the numerator of the NTF to zero, producing:

CI, C2 are not independent because they are related to Bl, 52 as specified by the NTF equation, previously described. The solution yields the four roots, as follows:

The structure shown in Figs. 2 and 3 allows two pairs of complex zeros, both of which have real parts equal to 1. This means they cannot be on the unit circle. However, if their angles are small enough, they will still provide enough attenuation. To actually be able to have zeros on the unit circle, more feedback loops (i.e., more coefficients) must be used.

Bl, 52 are selected so that preferably the zeros have the same angles as those required by the ideal transfer function. This is shown in Fig. 8, where the angles are exaggerated.

Bl, .52 are then selected to be negative, in which case the angle, α, of the respective zero is:

The values of Bl, B2 also depend on the values of K2 and K4. In general, the scaling coefficients k, shown in Fig. 2 as k^-k^ should be adjusted so noise shaper 802 is stable for the desired range of amplitudes for the input signals. Preferably, this is accomplished with the following criteria in mind:

• The scaling coefficients, k, are equal for the first and third integrators 822a (Fig. 2) and also for the second and fourth integrators 822b. This permits re-utilization of one hardware block 830 containing two integrators 822a and b and associated adders 848 without having to change scaling coefficients, k. Hardware block 830 is enclosed inside the dotted line in Fig. 2.

• The scaling coefficients, k, are only negative powers of two, so only hardwired shifts are used, without multiplication.

• The scaling coefficients, k, equalize the signal range at the integrator 822a and b outputs so the required word width is uniform throughout the structure.

• The scaling coefficients, k, set the stability range to be compatible with the desired input signal levels.

The scaling coefficients obtained for an input signal range of +/- 1.4 preferably, are: k, = 0.25 k₂ = 0.5 k₃ = 0.25 k_< = 0.5 k_a = 0.25 The feedback coefficient values Bl and B2, for positioning the zeros, are obtained using these scaling factors and preferably are:

Bl = -0.016470516 (quantized to 1/64(1+1/32) = -0.01611328) J32 = -0.002823487 (quantized to 1/256(1-1/4) = -0.002929688) The coefficients for denominator D in the NTF equation, H_E(z), above, are obtained by matching the terms in equal powers of z in the equation:

with the denominator D of the discrete filter to obtain the Wi values, shown above, and then, working through the equations given, together with the values of Bl and B2. In this embodiment, for Fig. 20, a unique solution exists. The preferred feedback coefficients A_rA_<, for positioning the poles, are:

A, = -4.724470536 A, = -4.204171137 A₃ = -4.04452877 A₄ = -0.837991461 These feedback coefficients can be quantized to 10 bits, before the STF begins to be affected inside the signal band, where:

A, = -4.71875 A, = -4.203125 A₃ = -4.03125 A, = -0.828125

The actual NTF magnitude is compared in Fig. 9 with the magnitude of a NTF obtained placing all the zeros at DC (z = 1). It can be seen that the noise power in the signal band is about 11.7 dB less in the selected structure, using Chebyshev zeros, than it is in the simpler one with all zeros at DC.

Signal Transfer Function (STF) For Noise Shaper Once the feedback coefficients, A; B; shown in Fig. 2 have been determined, the STF for noise shaper 802 is fixed. If the oversampling ratio is large enough, the STF will have little effect inside the signal band. Otherwise, the poles can be tweaked to some extent, but this is not desirable, because stability may be compromised. A better embodiment is to compensate for any distortion in the interpolation block 800 (Fig. 1). The magnitudes of the STF and the NTF are shown in Fig. 10 over the entire frequency range. The preferred STF response in the passband appears in more detail in Fig. 11. The group delay inside the passband is shown in Fig. 12. The passband tilt, if significant enough to violate the preferred ripple requirement for the entire playback path, can be compensated, preferably in interpolator 800. With regard to group delay distortion, however, it is still acceptable.

The difference between maximum and minimum group delay values is about 111.6 ns. The phase deviation from linear at 3.6 KHz with f, = 8 KHz is equal to:

W

The foregoing disclosure and description of the invention are illustrative and explanatory of the preferred embodiments, and changes in the size, shape, materials and individual components, elements, connections and construction may be made without departing from the spirit of the invention.

What is claimed is:

Claims

1. A 1-bit noise shaper circuit, comprising: a fourth order sigma-delta modulator network having a plurality of feedback paths, wherein said sigma-delta modulator network includes a plurality of integrators, wherein each said integrator includes an input, wherein said input of each said integrator is connected to a separate one of said feedback paths, and wherein a multi-bit digital signal input to said noise shaper circuit is converted to a 1-bit digital output signal.

2. The 1-bit noise shaper circuit of claim 1, wherein said circuit has a signal transfer function of:

» where X(z) is the digital audio input signal,

wherein

W₄=-1₄K_A+A₂B,K₄₊BιB₂K₄, wherein A_1-4 are pole positioning feedback coefficients, B₁.₂ are zero positioning feedback coefficients, and K_M are scaling factors.

3. The 1-bit noise shaper circuit of claim 1, wherein said 1-bit output signal includes a phase variation of no greater than about 0.011 degrees.

4. The 1-bit noise shaper circuit of claim 1, wherein said circuit includes a stopband noise edge of about fJ2 when the sampling rate of said multi-bit digital input signal is f..

5. The 1-bit noise shaper circuit of claim 1, wherein said circuit includes a stopband noise edge of about 4 KHz when the sampling frequency of said multi-bit digital input signal is about 8 KHz, and a stopband noise edge of about 24 KHz when the sampling frequency of said multi-bit digital input signal is about 48 KHz.

6. The 1-bit noise shaper circuit of claim 1, whereby said circuit includes the following discrete zeros and poles:

and

where, sz_k and sp_k are zeros and poles, respectively, of a continuous-time transfer function, and where

and

7. The 1-bit noise shaper circuit of claim 2, wherein scaling factors K_lw, and feedback coefficients Bj.₂ have the following values: k₁ = 0.25, k, = 0.5, k₃ = 0.25, and k₄ = 0.5; and B₁ = -0.016470516, B₂ = -0.002823487.

8. A noise shaping circuit of the type having a multi-bit digital input signal and a 1-bit output signal, comprising:

(a) a first summing node, having a plurality of inputs and an output, wherein one input of said first summing node is connected to said multi-bit digital input signal and another input of said first summing node is connected to a first feedback signal path;

(b) a first integrator, having an output and having an input connected to said first smnming node output;

(c) a second summing node, having a plurality of inputs and an output, wherein said output of said first integrator is connected to one said input of said second summing node and a second feedback signal path is connected to another said input of said second summing node;

(d) a second integrator, having an input and an output, wherein said input is connected to said output of said second summing node;

(e) a third summing node, having a plurality of inputs and an output, wherein said output of said second integrator is connected to one of said inputs of said third summing node and another of said inputs of said third summing node is connected to a third feedback signal path;

(f) a third integrator, having an input and an output, wherein said input of said third integrator is connected to said output of said third summing node;

(g) a fourth summing node, having a plurality of inputs and an output, whereby one input of said fourth summing node is connected to said output of said third integrator and another input of said fourth summing node is connected to a fourth feedback signal path;

(h) a fourth integrator, having an input and an output, wherein said input of said fourth integrator is connected to said output of said fourth summing node; and (i) a quantizer, having an input and an output, wherein said input is connected to said output of said fourth integrator, and wherein a 1-bit digital output signal is provided at said quantizer output.

9. A method of converting a multi-bit digital input signal to a 1-bit digital output signal, comprising the steps of:

(a) providing a multi-bit digital input signal to the input of a fourth order sigma-delta modulator, wherein said fourth order modulator includes a plurality of integrators, wherein each said integrator includes an input, wherein said fourth order modulator further includes a plurality of feedback paths, and wherein said input of each said integrator is connected to a separate one of said feedback paths; and

(b) converting said multi-bit digital input signal to a 1-bit digital output signal using said fourth order modulator.

10. A 1-bit noise shaper circuit for a digital-to-analog conversion circuit, comprising: a sigma-delta modulator network including a quantizer and a plurality of feedback paths, wherein each such feedback path provides a 1-bit feedback signal which is output from said quantizer, wherein said modulator network further includes a plurality of integrators, wherein each said integrator has an input, wherein the input of each said integrator is connected to a separate one of said feedback paths, and wherein a plurality of multi-bit digital signals are input to said noise shaper circuit and are converted to a respective 1-bit digital output signal.

11. The 1-bit noise shaper circuit of claim 10, wherein said sigma- delta modulator comprises a fourth order network.