IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II
1
Noise Weighting in the Design of ΔΣ Modulators (with a Psychoacoustic Coder as an Example)
arXiv:1309.6151v1 [cs.IT] 24 Sep 2013
Sergio Callegari, Senior Member, IEEE and Federico Bizzarri, Member, IEEE Abstract—A design flow for ΔΣ modulators is illustrated, allowing quantization noise to be shaped according to an arbitrary weighting profile. Being based on FIR NTFs, possibly with high order, the flow is best suited for digital architectures. The work builds on a recent proposal where the modulator is matched to the reconstruction filter, showing that this type of optimization can benefit a wide range of applications where noise (including in-band noise) is known to have a different impact at different frequencies. The design of a multiband modulator, a modulator avoiding DC noise, and an audio modulator capable of distributing quantization artifacts according to a psychoacoustic model are discussed as examples. A software toolbox is provided as a general design aid and to replicate the proposed results.
u(nT )
x nT
FF z
Linearized quantizer model nT
FB z
quantizer c
Figure 1. General block diagram of a discrete-time ΔΣ modulator and approximated linear model.
ideal filter to separate away the quantization noise would not be realistically available [6]. Thus, the designer should select I. I NTRODUCTION the best possible (good enough) NTF given actual conditions. While often associated to A/D conversion, ΔΣ modulators are In this paper, a design strategy for the NTF is proposed in fact coders that enjoy a wide range of applications. Generally allowing the quantization noise (including the residual in-band speaking, they exploit a nonlinear feedback architecture (Fig. 1) noise) to be shaped according to an arbitrary weighting (cost) to translate an analog or high-resolution digital input into a low- profile. This is not possible with conventional flows [5], [7], [8], resolution high-sample-rate digital signal with minimal loss which merely distinguish between the signal band and out of of fidelity [1]. As a matter of fact, their prevalent commercial band frequencies, aiming just at concentrating noise in the latter. deployment is as digital units, namely Digital ΔΣ Modulators While in principle valid for analog modulators too, the proposed (DΔΣMs) [2], used in tasks such as D/A and D/D conversion strategy is best suited for DΔΣMs since it delivers a Finite [1], fractional Phase Locked Loops (PLLs) [3], etc. Impulse Response (FIR) NTF, which may easily require a high A key phase in their design is the selection of the Noise order. The work builds on [6] that, owing to the interpretation Transfer Function (NTF) [4], fundamental to the preservation of DΔΣMs as heuristic optimizers for filtered-approximation of information content. This is particularly true for DΔΣMs, problems [9], [10], suggests that modulators can be matched whose digital nature frees the designer from many limitations. to their output/reconstruction filters (as exemplified by the However, the Literature is mostly concerned with analog applications in [11], [12]). Here, we introduce a change in architectures and many of its NTF considerations are over perspective illustrating how the same type of optimization can constrained or not directly applicable to DΔΣMs [2]. also benefit a wide range of applications where an explicit, Recall that the NTF derives from a linear approximation tangible filter does not exist but nevertheless there are known replacing the modulator quantizer with the superposition of a reasons for suffering more from some types of noise than from noise signal ((nT ) in Fig. 1, where T is the sample period). others. To do so, we emphasize by an alternative but equivalent Together with the Classical Model of Quantization (CMQ) mathematical derivation that even when a filter can be identified [5], this lets the modulator behavior be expressed by two only its magnitude response determines the NTF design, so items: the Signal Transfer Function (STF), from input u(nT ) that the response can be reinterpreted as a weighting. to output x(nT ), and the NTF, from (nT ) to x(nT ). In The applicability of the proposed design flow is wide: (i) principle, full preservation of information is possible if u(nT ), using on-off weighting functions (1 in the signal band and 0 passed through the STF, is decoupled in band (i.e., separable out of it) it reproduces conventional design methods; (ii) it by a linear filter) from (nT ) passed through the NTF. In deals with situations that cannot be easily managed otherwise, practice, a full decoupling is never possible. Even if it were, an such as multiband signals; (iii) most important, it simplifies the management of the sub-ideal behaviors that characterize This is a post-print version of a paper appearing in the IEEE Transaction on Circuits and Systems - Part II. Available through DOI http://dx.doi.org/10. real world applications, letting the residual in-band noise be 1109/TCSII.2013.2281892. Always cite as the published version. concentrated where it can be less harmful and the out-of-band Copyright © 2013 IEEE. Personal use of this material is permitted. However, noise be concentrated where the removal can be more efficient. permission to use this material for any other purposes must be obtained from The paper provides many application examples including the IEEE by sending an email to
[email protected]. S. Callegari is with the Advanced Research Center on Electronic Systems the design of a coder for audio applications distributing the for Information and Communication Technologies “E. De Castro” (ARCES) residual in-band noise according to a standard psychoacoustic at the University of Bologna, Italy. E-mail:
[email protected]. F. Bizzarri is with the Dipartimento di Elettronica, Informazione e Bioingeg- model [13]. Directions are also provided to download open neria at Politecnico di Milano, Italy. E-mail:
[email protected]. source code meant for the replication of the results in the
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II
2
(ii) even if the whole of the quantization noise could be pushed out of the signal band, it would be impossible to ignore II. BACKGROUND it, since no real world filter would be able to fully remove it. When a modulator such as that in Fig. 1 is linearized, the Furthermore, conventional strategies typically assume that B is relationships between NTF (z), STF (z) and the loop filters an interval and cannot be easily extended to multi-band cases. One can immediately see that the way in which the residual FF (z), FB (z) are in band quantization noise distributes can matter and that the ( ( STF (z) 1 NTF (z) = 1+cFF (z)FB(z) FF (z) = cNTF way in which the noise pushed out of band gets distributed is (z) (1) cFF (z) 1−NTF (z) important as well. Actually, these considerations have been clear STF (z) = 1+cFF FB (z) = (z)FB(z) STF (z) for a long time. Already in 1997, [15] attempted at designing where c = 1 is customarily assumed. This means that the NTF ΔΣ modulators for audio applications capable to distribute the selection fully determines the modulator whenever the STF residual in-band quantization noise so that it could be minimally is pre-assigned. Typically, specifications want STF (z) to be audible. However, the refinement of these older attempts is unitary or at most an integer delay z −d with d ∈ . In the somehow limited. More recently, [16] proposed a formal following, STF (z) shall be assumed to be 1 for simplicity. method to minimize the peak values of the in-band residual CMQ states that uniform quantization can be approximately noise (namely to minimize the maxima of NTF ei2πf ). modeled as the superimposition of noise, white in spectrum, Furthermore, [6] proposed an output filter aware design strategy independent from the quantized signal and uniformly distributed where the modulator NTF is matched to the filter in charge of within [−∆/2, +∆/2], where ∆ is the quantization step. It holds removing the out-of-band noise. relatively well whenever the modulator input signal is “busy” Here, we propose the introduction of a cost factor ν based 2 [5]. With this, the input noise power is σ2 = ∆ /12, and on a weighting function w : [0, 1/2] → + its Power Spectral Density (PSD) is uniform and equal to Z 12 2 Ψ (f ) = ∆ /6, with f normalized in [0, 1/2]. Consequently, the ν= Ψn (f )w(f ) df (4) noise component at the modulator output has a PSD 0 so that the NTF design can be based on its minimization. This 2 ∆2 Ψn (f ) = NTF ei2πf . (2) lets points (i) and (ii) be flexibly managed. Furthermore, it 6 obviously represents a generalization of conventional design The NTF choice is subject to some constraints. First of all, strategies as well as specialized ones like [15], since w(f ) can the modulator loop cannot be algebraic. Thus, cFF (z)FB (z) = be taken on-off or shaped according to given profiles such (1−NTF (z))/NTF (z) must include some delay. This condition as equal-loudness ones [13], [17]. With respect to [6] it can can only be satisfied if the NTF impulse response has a unitary be viewed as change in perspective since the design strategy zero lag coefficient. Secondly, one must guarantee that the becomes aware of a cost function, which may not have a modulator loop is stable. This condition is hard to tackle since corresponding block in the physical architecture of the system. it is not sufficient to look at the approximated linear model. To deal with the minimization, the NTF needs to be restricted However, it is known that a frequent mechanism that can break to a FIR form, to guarantee that non-convex expressions, the modulator operation is the overloading of the quantizer which would be unmanageable, are avoided [6]. A P order [5]. Thus, a common approach to favor stability consists in NTF (leading to a P order modulator) is described by P + 1 limiting the peak gain of the NTF, taking coefficients a0 , . . . , aP as in i2πf P kNTF k∞ = max1 NTF e
