The sound of software
Paul Schreier explores how scientific software has allowed some remarkable advances in the area of acoustics, which in turn have led to some very innovative products
The fundamentals of acoustics engineering have changed little over the decades. For instance, at its heart, a speaker still consists of a magnet, coil and some sort of diaphragm. Scientific software, however, has allowed researchers in acoustics to make some interesting discoveries and come up with innovative products and techniques. This article looks at how software has contributed to quite unusual speaker designs, noise-cancelling headphones, the noise analysis of rooms such as open offices and turbine halls, and how a psychoacoustic model of human hearing led to software in the latest consumer electronics that automatically adjusts levels of unwanted noise (such as adverts) while leaving desirable signals untouched.
Use any surface as a speaker
While we traditionally think of speakers as separate boxes, even if they are getting smaller and smaller, SFX Technologies (Dunfermline, Scotland) is using modelling software to help them design a new type of loudspeaker driver that uses virtually any surface – from a tabletop to walls, mirrors, dashboards, billboards or even bus shelters – as the loudspeaker. Users mount a Gel Audio transducer, which can be as small as 11 x 16 x 2.5mm, in permanent contact with a panel without being bonded to it. It can be attached, for example, with two-sided adhesive tape. The transducer’s magnet and coil receive analogue audio signals from an amplifier, and a slice of gel material transfers the acoustic waves to the panel. The result is a ‘speaker’ with good high-frequency response, but the big advantage is that this small driver generates very good bass response in the low frequencies without a large external speaker box.
Simulation is necessary to size the various mechanical components, such as the coil and magnet, and examine their effects. Further, while those two components produce unidirectional movement, the panel on the other side of the gel can produce a very complex waveform, especially at low frequencies consisting of lots of movement that might lead to distortion. This movement can also bend the mechanical driver and coil and lead to further distortion. So one task of the modelling is to find the right amount of gel and the best way to attach it to the surface. Too much gel makes the driver inefficient and nonresponsive; too little leads to distortion.
A Gel Audio driver (left) placed inside a television can produce very good bass response using the TV itself as the speaker surface. The Comsol simulation on the right illustrates and measures the SPL created by placing an SFX driver against the case of a TV set and the sound pressure level (SPL) at 1900Hz across a quarter of it.
Without modelling, says technical director Jordi Munoz, it would take three months to get to an initial prototype of a new design whereas with simulation, which is now being done with Comsol Multiphysics and the Acoustics Module add-on, a designer can achieve a first prototype in one month. This time difference is vital for this new technology where the company wants to get its devices on the market quickly, especially as the company expands from the commercial arena into consumer products.
Gel Audio drivers were initially popular in places where you don’t want speakers to be visible or accessible, such as in a bus stop or public-address systems in high-crime areas where, because the driver is behind a panel and no holes are necessary to let the sound through, there is little chance of vandalism. Now, though, Gel Audio drivers are being put into products such as small TVs so you can get good bass response, even without a subwoofer. The company is also looking at putting them into mobile phones for better audio response. Here it is very difficult to prototype a physical system, because handset manufacturers don’t want to hold up their development efforts to see how well such an added device would work in their phones. Simulation results can show handset developers what they can expect when the drivers are later incorporated into a phone so development can proceed on both the phone and the driver in parallel.
Block out annoying sounds
A development in acoustics that has captured the public’s attention is the noise-cancelling headphone, often used in airplanes or public-transit systems to eliminate ambient noise. While most of the world has gone digital, analogue circuits are still the primary technology here, explains Axel Grell, senior acoustical engineer at Sennheiser. Such headphones could be built with digital electronics, and this might soon be the case in professional versions such as in pilot’s headphones where power comes through the cable. But, for consumer models that run on batteries, the fast analogue/digital and digital/analogue converters that would be required consume a great deal of power, and the digital signal processor that performs the calculations also contributes to the power budget. Large, heavy batteries are simply not an option in portable consumer headphones. Grell notes that Sennheiser’s PXC 450 travel headphones can run for 20 to 30 hours on one AAA cell.
The PXC 450 noise-cancelling headphones from Sennheiser combine physical and electronic acoustic filtering to create flat response across the audio spectrum.
It’s also interesting to note that active noise-cancelling technology such as Sennheiser’s NoiseGard is not used throughout the audio spectrum; in this case only to roughly 1kHz. To achieve a flat frequency response in other regions of the audio spectrum, the headphones rely on passive damping of high frequencies along with high-quality diaphragms and speakers. While total noise reduction is possible at low frequencies or at individual high frequencies, top products such as the PXC 450 reduce noise by 90 per cent – achieving 100 per cent noise reduction today is almost impossible. That is because of the phase delay introduced between the loudspeaker that plays back the cancelling signal and the internal microphone that detects the remaining noise at the ear. This phase gap increases with frequency and could even lead to distracting feedback.
Design software has proven essential in developing such products. For its high-end products, Sennheiser uses diaphragms, speakers and microphones of its own design and manufacture. For this purpose, it turns to Ansys for structural and acoustic analysis. For instance, with a good transducer, the eigenforms are located at high frequencies and don’t disturb the phase response, and Ansys makes it possible to find the eigenforms for a given geometry. The designers have validated their simulation results with good agreement to empirical data from laser Doppler interferometers that visually show a diaphragm’s operation.
Going a step further, the engineers next use Ansys to perform structural and acoustic modelling on the entire headphone system, including the mechanical acoustic filtering it performs, and determine which part of the mechanical frequency response must be enhanced with analogue amplifiers and filters. With this knowledge, they create a mathematical model of the acoustic response in Matlab, where they design the transfer functions of the analogue filters that produce the flat frequency response across the entire audio range. A final step is to take the theoretical filter information and work with the MicroCap filter design software to determine the values of the resistors and capacitors that work with op-amps to implement the desired amplification for the flat response.
Analyse all types of spaces
The analysis of room acoustics has had a long history in concert halls and performance areas, but those techniques are being expanded to cover other applications such as in open-plan offices to see how loud voices in one area are in other areas, in underground stations to make sure that announcements are understandable, or in turbine halls to simulate noise levels and to find out which measures need to be taken in order to keep noise levels within the acceptable (making the machinery more quiet, encapsulating the machinery, installing absorbent materials, etc.).
These spaces are often irregularly shaped and have large or multiple sources of sound. Researchers at Odeon A/S in Denmark believe that the best approach to handle them, including all possible echoes and reflections, is to use a hybrid approach that combines two things: first, the image source method (which treats a wall like a mirror and where the wall can be removed and modelled by another source the same distance from the wall as the original source was); second is the ray-tracing method (which treats sound like rays of light, which lose energy when they reflect off surfaces, and to find the response at a point in space you trace straight line paths from the receiver to the source).
In this OpenGL rendering of a model of a turbine hall, the colours map the acoustic reflectance of the surfaces onto a RGB value: red indicates low-frequency reflectance; green, mid-frequency reflectance; and blue, high-frequency reflectance. The grey scales indicate a flat frequency response of the colours, e.g. black is 100 per cent absorption.
Using this approach, the Odeon software package, which began development in 1984 at the Technical University of Denmark, can handle a wide variety of indoor situations. For instance, the industrial version allows the definition of line and surface sources. To see how it works, consider a study of a turbine hall at a power plant. The first task is to create a 3D model of the room geometry. Although the room is large (153 x 34.5 x 20m) and contains two turbines, the model is quite simple and can omit many details because the wavelength of the sound of interest is typically less than a metre; 1,000Hz (wavelength 0.34m) is often considered mid frequency, and the 500,1000,2000Hz octaves are considered very important for speech.
Next, because the noise sources (the turbines) are big and complicated, each was modelled with 17 surface sources and two point sources that account for ball bearings.
A grid mapping of the calculated A-weighted sound pressure levels in a turbine hall done with Odeon. The two turbines are modelled as four point sources (ball bearings) and 46 surface sources.
One very useful feature of a room acoustic computer model is the ability to calculate the sound pressure at a large number of receiver positions and display the result in a grid map. In an interesting enhancement to older acoustics packages, Odeon can also perform auralisation, whereby it can make simulation results audible. This has proven an efficient tool, for instance, in the study of speech privacy in open-plan offices. You can put on a pair of headphones and hear what someone would hear from a certain sound source if they were standing at a given point in the office. The input to the simulation is a recording of the sound signal free of any reflections and reverberation, and the software then modifies it according to the paths it would take and the echoes and the surfaces reflecting it (partly absorbing or hard reflective).
Clear listening even at low volumes
Much of our acoustic experience takes place behind the scenes before the signals reach the loudspeaker, and here Dolby Laboratories has established itself as a leader in audio signal processing. One of its latest developments is Dolby Volume, which solves the problem of volume inconsistencies among channels and sources – for example, between a TV programme and adverts or when switching from channel to channel – to deliver a consistent volume level to the viewer. In addition, even at low volumes, as when watching TV late at night, this technology can adjust the spectral balance so that dialogue remains clear and loud effects or music passages retain their impact. Harman Kardon was the first company to announce the inclusion of Dolby Volume in an audio/video receiver in its Model AVR 755. In addition, this technology has been integrated into Toshiba’s new Regza ZH500 and ZV500 series LCD HDTVs, which were launched in Japan last year.
Traditional methods for automatic volume control change the gain globally based on a simple electrical measure of an audio signal’s level. This approach, however, can often adjust the audio at an inappropriate time, for instance, when artificially levelling the loudness of what should be a decaying piano note. These conventional methods can give the audio an unnatural boost, known as ‘breathing’, whereby all low-level sounds and noise are amplified along with the desired sound. The opposite effect is ‘pumping’ or lowering a loud sound, thereby taking the life or intelligibility out of the audio.
Dolby Volume handles the job very differently, and its uniqueness lies in the fact that for the first time the measurement, analysis and control of the volume level of audio is performed using a perceptual processing engine built on an advanced psychoacoustic model of human hearing; it continuously ‘listens’ to the audio as a human does to accurately predict when a soft or loud sound is coming. It also uses auditory scene analysis to detect how much processing to apply to a given event such as the decaying piano note. It breaks the audio down into multiple frequency bands, operating much like the human ear does, and adjusts the volume in each band based on a nonlinear model of loudness growth. To arrive at these volume adjustments, the software performs a variety of analyses, looking at how the audio changes over time in each frequency band and segmenting the audio based on these changes.
A graphical analysis of Dolby Volume levelling incoming audio.
To develop Dolby Volume, notes senior staff engineer Alan Seefeldt, the R&D department reaches first for Matlab, which is their workhorse for audio development thanks to its extensive library of signal-processing functions. However, when testing and refining the prototype algorithms, the team needs to adjust more than a hundred different parameters in the software in real time and hear the results, and this is something Matlab does not handle very easily. So the designers create another prototype in C or C++ along with OpenGL graphics to display what the software is doing. In the figure below, which shows left/right stereo channels, the yellow bars represent the perceived loudness spectrum of the left and right channels of the incoming audio as measured by Dolby Volume’s psychoacoustic model. The blue lines represent the gains being applied by Dolby Volume to each band, and the more these lines diverge, the greater the inconsistencies in volume.
To stream audio through the prototype software, they also make use of ASIO, a proprietary audio interface standard developed by the German company Steinberg that bypasses Window’s mixing kernel and thus provides low-latency direct communication between computer audio software and hardware.
A big challenge, notes Seefeldt, is listening to enough content to ensure that the algorithms are optimised. Here it is crucial that he be able to twiddle the various parameters in real time and hear the results. Some parameters can be preset into reasonable ranges based on psychoacoustic studies and then just fine-tuned, but others need more work. He also notes that visual feedback of what the algorithm is doing is valuable. The audio can sound fine 90 per cent of the time and then come cases where things go haywire, and with a visual indication he can identify the error more quickly. ‘There are certain things you might not hear, but you know how you want the parameters to move over time, and here visualisation is key,’ adds Seefeldt. Once the parameters are all locked in, the user controls are then boiled down to simplify operation for the consumer, in this case to an on/off button and a levelling control knob.
Acoustic modelling is now a very sophisticated business, aided in no small part by the computational power offered by software packages, resulting in better products for consumers.