Constraint Optimized Multichannel Mixer-limiter Design
1 st Yuancheng Luo, 2 nd Dmitriy Yamkovoy, 3 rd Guillermo Garcia · 2025
Read Constraint Optimized Multichannel Mixer-limiter Design by 1 st Yuancheng Luo, 2 nd Dmitriy Yamkovoy, 3 rd Guillermo Garcia (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-06769.txt.
Introduction
Yuancheng Luo, Dmitriy Yamkovoy, and Guillermo Garcia design a combined audio mixer and limiter for loudspeaker arrays in a 2025 paper. Mixers route and blend channels while limiters cap loudness; the two are usually designed separately because joint design is computationally heavy. Separate designs can fight each other and add distortion.
Their coupled design is posed as a quadratic optimization with linear constraints over the channel gains. New methods trim the variables and approximate the goal cheaply enough for real-time use. Window design for overlapping audio frames also improves. Experiments show less distortion than the split approach, with clear trade-offs in computing cost that engineers can tune.
Catalog record
| CREATOR | 1 st Yuancheng Luo, 2 nd Dmitriy Yamkovoy, 3 rd Guillermo Garcia |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-06769
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Constraint Optimized Multichannel Mixer-limiter Design 1 st Yuancheng Luo, 2 nd Dmitriy Yamkovoy, 3 rd Guillermo Garcia Abstract Multichannel audio mixer and limiter designs are conventionally decoupled for content reproduction over loudspeaker arrays due to high computational complexity and run-time costs. We propose a coupled mixer-limiter-envelope design formulated as an efficient linear-constrained quadratic program that minimizes a distortion objective over multichannel gain variables subject to sample mixture constraints. Novel methods for asymmetric constant overlap-add window optimization, objective function approximation, variable and constraint reduction are presented. Experiments demonstrate distortion reduction of the coupled design, and computational trade-offs required for efficient real-time processing. I Introduction Multichannel audio content reproduction over loudspeaker arrays has grown in popularity in recent years with the proliferation of low-cost sound-bar and smart-speaker consumer electronics. Such audio-reproduction systems are typically resource-constrained compared to professional-grade loudspeakers in terms of available transducer-level digital-electrical headroom, and acoustic output. A conventional method digitally compresses the audio content’s dynamic-range throughout channel-mixing stages to stay under digital full-scale levels, maximize loudness, and satisfy studio standards [ 1 , 2 , 3 , 4 ] . A channel matrix-mixer [ 5 , 6 ] or matrix-decoder [ 7 , 8 , 9 ] therefore allocates headroom between input channels and output transducers; dynamic-range controllers (DRCs) [ 10 , 11 , 12 ] such as peak-limiters are thereby placed downstream for transducer and amplifier protection. Several deficiencies of the conventional method are known: A multichannel matrix-mixer can conservatively pre-allocate headroom for each input channel such that maximum channel mixture levels minimally activate downstream DRCs. Pre-allocation however distorts the audio mixture in the absence of run-time monitoring; a channel’s upper dynamic range may be unnecessarily lowered when other channels are sparse, and the choice of mixing gains can alter the spectral and channel balance of the original content. Terminal DRCs operating on mixtures of channels per transducer can limit at different times, intermittently distorting both channel balance and the loudspeaker array’s directivity. We address both pre-allocation and terminal-limiter problems by coupling time-varying channel-gain reduction with per-sample constraints of the channel mixtures via a sequence of quadratic programming (QP) problems. Section II presents our QP mixer-limiter design, and relates the QP’s feasibility with a novel constant overlap-add (COLA) [ 13 ] constrained gain envelope construction. Section III presents the channel-mixture’s distortion objective and derives the QP objective from the former’s optimal Taylor series approximation. Section IV extends the QP formulation to joint multi-band [ 14 ] multi-content mixers, and introduces novel variable and constraint size reduction methods for efficient computation. Section V shows experimental results for distortion reduction and computational performance. II Quadratic Program Mixer-Limiter Let \boldsymbol{S}\in\mathbb{R}^{F\times N} be a matrix of N input channel column-vectors of F samples in an audio frame. Each of the N input channels are mixed with independent variable gains in vector \boldsymbol{x}=\left[{x_{1},\ldots,x_{N}}\right]^{T} to produce a single output channel mixture limited in dynamic range and satisfying the equivalent modulus linear constraints given by \begin{split}&-\tau\leq\sum_{n=1}^{N}S_{mn}x_{n}\leq\tau,\quad\tau\geq 0,\quad 1\leq m\leq F,\\ \end{split} (1) where \tau is a user-specified non-negative threshold for any mixture of samples within a frame. The variable gains are subject to non-negative bounds under unity and are equivalent to box-constraints given by \begin{split}&0\leq l_{n}\leq x_{n}\leq u_{n}\leq 1,\quad 1\leq n\leq N,\\ \end{split} (2) and therefore only apply gain reduction to each input channel. The gain variables \boldsymbol{x} , constrained to the feasible space that satisfy the linear constraints, are then found by minimization over a quadratic polynomial objective function. The latter’s standard form is given by \begin{split}f(\boldsymbol{x})=\frac{1}{2}\boldsymbol{x}^{T}\boldsymbol{Q}\boldsymbol{x}+\boldsymbol{c}^{T}\boldsymbol{x}+d,\end{split} (3) where symmetric matrix \boldsymbol{Q}\in\mathbb{R}^{N\times N} , vector \boldsymbol{c}\in\mathbb{R}^{N\times 1} , and constant d parameterize our distortion objective in section III . Minimizing ( 3 ) subject to constraints ( 1 ), ( 2 ) in vectorized form is given by \begin{split}\boldsymbol{x}_{*}&=\arg\min_{\boldsymbol{x}}f(\boldsymbol{x}),\quad\textrm{s.t. }-\boldsymbol{\tau}\leq\boldsymbol{S}\boldsymbol{x}\leq\boldsymbol{\tau},\quad\boldsymbol{l}\leq\boldsymbol{x}\leq\boldsymbol{u}\leq\boldsymbol{1},\end{split} (4) where \boldsymbol{\tau}=\tau\boldsymbol{1}=\tau\sum_{i=1}^{F}\boldsymbol{e}_{i}\in\mathbb{R}^{F\times 1} is the vector of constant threshold \tau over frame-samples, \boldsymbol{e}_{i} is the standard basis, and \boldsymbol{l},\boldsymbol{u}\in\mathbb{R}^{N\times 1} are the vectors of lower and upper bound respectively. We now construct the limiter’s gain envelope across frame-wise solutions to ( 4 ). Constrained Limiter-Envelope Design: Attack, hold, and release-time parameters are commonly used to restrict the velocity and shape of a limiter’s gain envelope, constraining the latter to be smooth and reducing any audible distortion when multiplied by the input signals [ 15 ] ; attack refers to the early portion of the gain envelope of increasing gain reduction, hold is the middle portion of constant gain reduction, and release is the late portion of decreasing gain reduction. We can define an envelope function with attack, hold, release dynamics, and supports over the solutions to ( 4 ) across overlapped audio frames with look-ahead. Given an audio stream Y(m,n) of N input channels for sample index m and channel index n , let \boldsymbol{S}^{\left\{{k}\right\}}\in\mathbb{R}^{\bar{F}\times N} be the k^{th} audio frame of size F augmented with L look-ahead samples per channel, where \bar{F}=F+L , be defined as follows: \begin{split}S^{\left\{{k}\right\}}_{mn}=\left\{\begin{array}[]{cc}Y(m+(k-1)F,\,n),&1\leq m\leq F+L\\ 0,&\textrm{otherwise}\end{array}\right..\end{split} (5) For the k^{th} augmented frame, define the QP minimizer \boldsymbol{x}_{*}^{\left\{{k}\right\}} s.t. the sample constraint set \boldsymbol{\xi}^{\left\{{k}\right\}} in the frame and look-ahead as follows: \begin{split}&\qquad\,\,\boldsymbol{x}_{*}^{\left\{{k}\right\}}=\arg\min_{\boldsymbol{x}}f(\boldsymbol{x})\quad\textrm{s.t. linear constraints }\boldsymbol{\xi}^{\left\{{k}\right\}},\\ \boldsymbol{\xi}^{\left\{{k}\right\}}&=\left\{\begin{array}[]{rl}-\boldsymbol{\tau}\leq\boldsymbol{S}^{\left\{{k}\right\}}\boldsymbol{x}\leq\boldsymbol{\tau},&2\bar{F}\textrm{ Mixture limits \eqref{EQ:QPL:CONSTR_SAMP}}\\ \boldsymbol{0}\leq\boldsymbol{x}\leq\boldsymbol{u}\leq\boldsymbol{1},&2N\textrm{ Variable bounds \eqref{EQ:QPL:CONSTR_GAIN_ABS}}\\ \end{array}\right.,\\ \end{split} (6) where \boldsymbol{S}^{\left\{{k}\right\}} replaces \boldsymbol{S} in ( 4 ), \boldsymbol{\tau}=\tau\boldsymbol{1}\in\mathbb{R}^{\bar{F}\times 1} , and the variable lower bounds are set to \boldsymbol{0} to ensure that a feasible solution exists. The enveloped output mixture y(t) at time t is therefore given by the sum of gain enveloped inputs and the envelope function v:\mathbb{R}\rightarrow\mathbb{R}^{0+} : \begin{split}y(t)=\sum_{n=1}^{N}Y(t,n)v_{n}(t),\quad v_{n}(t)=\sum_{k=0}^{\infty}W_{n}(t-kF)x_{*n}^{\left\{{k+1}\right\}},\end{split} (7) where v_{n}(t) are weighted mixtures of solutions across frames. We show the bounds -\tau\leq y(t)\leq\tau and 0\leq v_{n}(t)\leq 1 are satisfied via the design of a weighting function W_{n}(t) . Fig. 1: Blue and red envelopes v_{n}(t) are COLA weighted mixtures of solutions in ( 7 ), belonging to different input channels, and computed over augmented frames \boldsymbol{S}^{\left\{{k}\right\}} in ( 5 ) with different look-ahead times. Observe in Fig. 1 that the convex set of solutions from consecutive frames \alpha\boldsymbol{x}_{*}^{\left\{{k}\right\}}+(1-\alpha)\boldsymbol{x}_{*}^{\left\{{k\pm 1}\right\}},\,0\leq\alpha\leq 1 is non-empty for the half-space intersection of \boldsymbol{\xi}^{\left\{{k}\right\}}\cap\boldsymbol{\xi}^{\left\{{k\pm 1}\right\}} common mixture-limit constraints from overlapped look-ahead and frame samples of \boldsymbol{S}^{\left\{{k}\right\}},\boldsymbol{S}^{\left\{{k\pm 1}\right\}} respectively. The set of solutions v_{n}(t) in ( 7 ), evaluated across multiple consecutive frames, is convex and satisfies the mixture-limited constraints of ( 1 ) at time t if W_{n}(t) are window functions with bounded non-negative supports in the frame and look-ahead interval, zero-valued elsewhere, and have the COLA [ 13 ] property given by \begin{split}\left\{\begin{array}[]{rc}0\leq W_{n}(t)\leq 1,&1\leq t\leq F+L\\ W_{n}(t)=0,&\textrm{otherwise}\end{array}\right\},&\quad\textrm{Bounded supports},\\ \sum_{k=\minus\infty}^{\infty}W_{n}(t-kF)=1,\quad\forall t\in\mathbb{R},&\quad\textrm{COLA property}.\end{split} (8) For bounded time t , the gain envelopes \boldsymbol{v}(t)=\left[{v_{1}(t),\ldots,v_{N}(t)}\right]^{T} span Barycentric weighted \boldsymbol{x}_{*}^{\left\{{k}\right\}} of consecutive sequences of frame index k satisfying 1+(k-1)F\leq t\leq(k-1)F+\bar{F} . We therefore design window function W_{n}(t) with bounded COLA and dynamics constraints for smoothly transitioning between supports or frames as to minimize discontinuities in the enveloped output mixture y(t) . Dynamics Constrained COLA Window Design: We can approximate a smooth COLA window W(t) with characteristic attack, hold, and release dynamics via constrained optimization over the uniformly-sampled and integer-spaced window samples \omega(t)=W(t),\,t\in\mathbb{Z} . Let the characteristic dynamics of W(t) be defined by its first-derivative’s intervals w.r.t. attack-release onsets as follows: \begin{split}\begin{array}[]{cccc}\textrm{Attack}&{d{W}}/{d{t}}\geq 0&1\leq t<T_{A}&\textrm{Attack-onset }T_{A}\\ \textrm{Hold}&{d{W}}/{d{t}}=0&T_{A}\leq t<T_{R}&\textrm{Release-onset }T_{R}\\ \textrm{Release}&{d{W}}/{d{t}}\leq 0&T_{R}\leq t\leq M&\textrm{Window-size }M.\\ \end{array}\end{split} (9) The