A Low-power System for Audio Noise Suppression: a Cooperative Analog-digital Signal Processing Approach

In this paper, we introduce the concept of cooperative analog-digital signal processing, and its application in the development of a real–time noise suppression system. The algorithm implemented is designed to reduce stationary background noise while preserving the non–stationary signal component. The algorithm is based on digital signal processing foundations that are slightly adjusted for use in the continuous-time domain. Because the system relies on analog computation rather than digital, it has benefits such as extremely low power consumption and real-time computation. The analog circuit elements are based on new floating-gate circuits that are small, efficient, and programmable—making it possible to set and tune bias points, offsets, and filter parameters under digital control. 1. COOPERATIVE ANALOG-DIGITAL SIGNAL PROCESSING New advances in analog VLSI circuits have made it possible to perform operations that more closely reflect those done in DSP applications, or that are desired in future DSP applications [1, 2, 3, 4, 5, 6, 7]. Further, analog circuits and systems can be programmable, reconfigurable, adaptive, and at a density comparable to digital memories (for example, 100,000+ multipliers on a single chip) [8, 9, 10, 11, 5]. These properties have been almost exclusively associated with digital processing, but the addition of small, dense, programmable analog circuits provides a framework in which to create cooperative analog/digital signal processing systems that benefit from both approaches to make something better than the sum of its parts. We define cooperative analog-digital signal processing (CADSP) as the research field using combinations of programmable analog signal processing and digital signal processing techniques for real-world processing. Neither analog signal processing nor digital signal processing can exist in current technologies without the other; that is, realworld signals are analog while much of the modern control and communication is digital. Therefore, a primary question is where to partition the analog–digital boundry to enhance the overall functionality of a system by utilizing analog/digital computations in mutually beneficial way (Figure 1). CADSP allows more freedom of movement for the partition between the analog and digital computation. CADSP is a superset of mixed-signal research in that it focuses heavily on algorithms as well as circuit implementation. By adding functionality to our analog systems, we enhance the capabilities of the controlling digital system, and therefore, the entire product under consideration. A full discussion of this partition problem can and will encompass several research papers. The range of applications for these approaches reaches from auditory and speech processing, to beam-forming, multidimensional signal x(t) x(t) y(n) y(n) (a) Digital Signal Processing (b) CADSP A/D Converter DSP Processor ASP IC DSP Processor A/D Converter Fig. 1. Illustration of the tradeoffs in cooperative analog/digital signal processing. We assume the typical model of signals coming from real-world sensors, which are analog in nature, that need to be utilized by digital computers. The inverse problem, digital signals going to real-world actuators, is similar in nature. One approach is to put an analog-to-digital (A/D) converter as close to the sensor signals as possible, and allow the remainder of the computations to be performed digitally. An alternate approach is to perform some of the computations using analog signal processing, requiring simplier A/D converters, and reducing the computational load of resulting digital processors. One could group this analog computation and A/D converter as a specialized A/D converter that gives more refined information (Fourier coefficients, phonemes, etc.) than a literal map of the incoming signal. The question of where to put this boundry line strongly depends upon the particular requirements of an application. processing, and radar computations, communications processing, and image processing and recognition. The following sections briefly discuss CADSP advantages and disadvantages followed by a discussion of the development and implementation of a background noise suppression system based on the CADSP approach. 2. THE CASE FOR COMBINING ANALOG AND DIGITAL SIGNAL PROCESSING One might wonder why introduce analog signal processing at all, since the current framework of immediately digitizing incoming signals, illustrated as the top half in Fig. 1, seems to be working well in current practice. As more computational power is required to implement more digital functionality, the needs have largely been met by transistor scaling and the advantages flexibility in programming. However, current trends are, and have been, finding unique challenges that provide an opportunity for a new perspective. Several factors suggest using analog signal processing in conjunction with digital systems to meet these challenges: 1. power consumption / efficiency requirements, 2. A/D converter requirements, 3. size constraints, and 4. problem “fit.” The analog VLSI signal processing circuits are programmable and adaptive using floating–gate circuits. The programmable and adaptive aspects of floating–gate analog VLSI are significant and make practical a variety of signal processing applications. One system that holds promise for the implementation of a large class of signal processing application in analog VLSI is the field programmable analog array presented in [12]. Power consumption in DSP microprocessors, as measured in mW/MIPS, has been decreasing by half about every 18 months— a phenomena known as Gene’s law [13]. This trend of decreasing power consumption has been keeping pace with Moore’s law and has helped make possible the increasing proliferation of portable electronics. Even so, device functionality, and the amount of signal processing, is often primarily constrained by a fixed power budget. A custom analog approach can often achieve an increased efficiency ( = Bandwidth·Power) of a factor of 10 over a custom digital approach [14]. If Gene’s law continues to hold, digital systems will catch up to current analog efficiencies within about 20 years! Therefore, migrating some applications to analog processing could provide up to a 20 year jump in functionality relative to a purely digital roadmap. The greatest efficiency occurs when matching physics of problem to physics of silicon medium. This increase in efficiency can be used for reducing the power efficiency of a given problem, or addressing computational problems that are considered intractable by the current digital roadmap. Analog to digital (A/D) converters also can be a limiting factor in signal processing systems and A/D requirements are becoming an increasingly large part of the system design constraints. The system demands on many current systems require very high resolution / high performance A/D converters; the resulting A/D converter block is often consuming a large fraction of the power budget, as well as system design time. While digital processor efficiency may be increasing rapidly, partially as a result of transistor scaling; scaling is not helping as much for A/D converters. A/D converters have roughly been increasing resolution at 1.5 bits / 5 years at the same performance, and quickly running into additional physical limits which might further slow this progress. In many problems the need for enormous speed and resolution originates in recovering “low-information” signals over a wide dynamic range or in a noisy background (e.g. CMOS imaging, software radio). By utilizing analog signal processing at the front-end, we can significantly reduce the A/D converter complexity, and the overall system complexity. Circuit size constraints also favor analog VLSI circuits in many cases and it is often possible to take advantage of device physics to perform complex operations with only a very few transistors. For example, an analog multiplier can store the coefficient and perform the multiplication using only as many transistors as would be needed to store a four bit coefficient in digital memory. Besides multiplication there are many elemental operations that can be efficiently implemented in floating–gate analog VLSI. These include filters, adaptive multipliers, a large range of non-linear functions, decision circuits, and others (see also [5, 11]). 3. NOISE SUPPRESSION SYSTEM The recent prosperity of portable computing and communication devices has resulted in a renewed interest in audio signal enhancement by suppressing additive background noise from corrupted noisy signal. While most noise suppression methods focus on processing sampled audio signals, Because noise suppression is 1The A/D converter limitations on a power constrained systems is somewhat similar to the problems generated as processor speeds progressed much faster than memory speeds in recent years. BPF