19 th INTERNATIONAL CONGRESS ON ACOUSTICS MADRID, 2-7 SEPTEMBER 2007

Transcription

1 19 th INTERNATIONAL CONGRESS ON ACOUSTICS MADRID, 2-7 SEPTEMBER 2007 A MATLAB SIMULATION OF SHOEBOX ROOM ACOUSTICS AND ITS EDUCATIONAL USE IN A MUSIC TECHNOLOGY COURSE. PACS: Sv Campbell, Douglas R. School of Computing, University of Paisley, High Street, Paisley PA1 2BE, Scotland, U.K; d.r.campbell@paisley.ac.uk ABSTRACT A simulation of the acoustics of a simple rectangular prism room has been constructed using the MATLAB m-code programming language. The program Roomsim may be used as an educational tool for illustrating the image method of simulating room acoustics and various reverberation effects, but is powerful enough also to be used as a tool to generate signals for the audio signal processing research community. The program is menu driven and has been made freely available on the MATLAB user contributed programs website [1]. This paper describes aspects of the program, and an undergraduate student coursework exercise as an example of its use in teaching a Music Technology course module entitled Advanced Audio Signal Processing. INTRODUCTION Several versions exist, in various computer languages, of a FORTRAN routine for simulating the acoustics of a simple empty rectangular prism ( shoebox ) room first published by Allen and Berkley [2] in This image-source method is often used as a means of generating audio signals incorporating sufficiently realistic reverberation to demonstrate room acoustic effects and for the preliminary testing of audio processing algorithms. Commercially available programs for simulating architectural acoustics can be used for these purposes, but they are often considered inappropriate due to their complexity and high cost which reflects the degree of geometric and acoustic realism they deliver. Many signal processing researchers use the MATLAB technical computing language to develop their algorithms because of its ease of use, powerful library functions and convenient visualisation tools. A search for a Windows/Unix MATLAB implementation of shoebox room acoustics found none that incorporated the combination of features to satisfy requirements for quickly and easily generating monaural/stereo/binaural audio signals, sufficiently realistic for educational demonstration of the influence of room acoustic effects and for experimental research assessment of audio signal processing algorithms. Following a presentation at a Binaural Hearing workshop, Palomäki made available the image method m-code developed to satisfy the particular needs of the SPHEAR project [3]. This MATLAB m-code was modified by Campbell to form the core of the Roomsim program reported here. In its present version Roomsim includes many additional features and functions to support simulation of the acoustics of a shoebox room within which simulated directional sensors (e.g. microphones or human heads) can be placed to allow estimation of the audio signal received at the sensors/ears from a number of omni-directional audio sources placed within the simulated room. The program is intended to provide acoustic impulse responses and reverberant audio signals sufficiently realistic for demonstrating some issues of room acoustics and audio effects to undergraduate and postgraduate students, and for preliminary evaluation of audio processing algorithms prior to more realistic, and time consuming, experiments. Care has been taken to correctly implement the image source model of shoebox room acoustics, but it is not a simulation of enclosure acoustics at the level of architectural acoustic simulations e.g. CATT Acoustics ( nor is it a real time auralisation system e.g. SLAB ( 1

2 This paper consists of two main parts: an introduction to some features and facilities of the Roomsim program, and an example of its educational use. THE ROOMSIM PROGRAM The Roomsim program uses the Allen and Berkley image-source algorithm [2], extended to include frequency dependent sound absorption coefficients. It computes an acoustic impulse response from each omni-directional primary source to a receiver system. Materials for each of the six major surfaces of the simulated room may be selected from a list of 20 standard building materials (plus 4 test cases including anechoic) whose frequency dependent absorption coefficients have been tabulated. The receiver system may be single sensor (mono microphone), sensor pair (stereo microphone array) or simulated human head. The receiver system as a whole can be oriented in 3D space, as can the individual sensors of a sensor pair. The single and dual sensor receivers can be configured with directionally sensitive elements and Peterson s interpolation process [4] is incorporated. Several directional sensitivities are provided including standard microphone polar responses e.g. omni-directional, bidirectional, cardioid. Simulation of the acoustic properties of the human head utilises the Head Related Impulse Response (HRIR) data from measurements made on a KEMAR mannequin at the MIT Media Lab [5], and from 42 human subjects and a KEMAR at the Center for Image Processing and Integrated Computing (CIPIC), University of California, Davis [6]. Both of these organisations have kindly agreed to the distribution of their data within the Roomsim package. The m-code of Roomsim has been written to avoid the need for additional MATLAB toolboxes, to avoid functions that cannot successfully be compiled to produce stand-alone executable code, and to allow operation on Windows and UNIX/Linux platforms. Thus, only a basic MATLAB installation is required to run and develop the m-code, and users without a MATLAB installation may run the executable version of Roomsim on a Windows PC. The program was developed using MATLAB v 6.5 rev. 13 on a PC (1.5 GHz Intel Pentium4, 512 MB RDRAM, 30GB HDD UDMA 100 IDE, AGP Graphics) running Windows On this system (using an example from [2]) simulating a single source and single sensor in a room of size 3*4.6*3.8 m and RT 60 =0.32 s, the core calculation of 28,500 images took approximately 42.5 s i.e. 1.5 ms per image. It has been run on desktop and notebook PCs down to 200 MHz clock frequency with 128 MB RAM. Reflection orders >10, impulse responses >10,000 samples and multiple sources increase the computational load resulting in longer run times, especially of the core impulse response calculation and the 2D and 3D graphics. It has run successfully without modification on all releases of MATLAB up to and including the current R2007a (MATLAB 7.4). The development of Roomsim has been shaped by the intention that this implementation would be made freely available and open-source for educators and digital signal processing (DSP) researchers working in areas of audio signal processing, music technology, speech enhancement and hearing. The program has undergone extensive testing to verify its numerical correctness and has been released as free software, under the GNU General Public Licence, to encourage its dissemination and development. The m-code source files execute within MATLAB on an MS-Windows PC or a Unix/Linux platform. These source files, an HTML user guide, and a compiled stand-alone executable for non-matlab Windows installations have been posted on MATLAB Central File Exchange [1], from which the package has been downloaded more than 3000 times to date. An alternate download location is available via the University of Paisley [7]. Roomsim operation A core aim was to provide user interaction having a suitable balance between flexibility for researchers and ease of use for undergraduate students. The user specifies various parameters [Table 1], such as the dimensions of the room, its surface materials and the locations of the primary source(s) and the receiver system. This can be done interactively through the menu prompt system, or by submitting a Microsoft Excel spreadsheet, or a text file, or by selecting a MATLAB file saved from a previous run. The program performs various sanity checks and an estimate of the reverberation time (RT 60 ) of the simulated room is used to size the impulse response to be computed. The source to receiver responses are then computed using the method of images, modified to take account of the variation of surface absorption coefficient with frequency, path length (1/R effect), and frequency dependent attenuation due to air, if desired. If a simulated head has been selected the acoustic impulse from each image-source 2

3 direction is convolved with the relevant quantised angular HRIR data. The individual imagesource responses are then accumulated to form the complete pressure impulse response from each primary source to the receiver sensors and the results plotted and saved to file. The impulse response data can then be convolved, within the program, with the users own monophonic audio files (*.wav, *.au or *.mat format). The resulting audio data related to each simulated primary source are then summed to produce the combined acoustic signal at each sensor/ear e.g. to create a cocktail party of speech/music and noise signals from different locations with their appropriate simulated reverberation. The resulting monaural, stereo, or binaural audio data can be saved as an audio file and may be played within Roomsim or by using a compatible media player or audio editor. Table 1.- List of user selectable parameters and facilities (extract) User Configurable Parameters Humidity of air (modifies air absorption coefficient). Temperature of air (modifies the speed of sound). Sampling frequency (at present 8 khz to 44.1 khz). Enclosure dimensions (Lx, Ly, Lz). Surface absorption for each of the room s six surfaces (at present 24 frequency dependent surface types). Air absorption model may be ON or OFF. Receiver coordinates (x, y, z). Receiver type (single sensor, sensor pair, HRIR). Sensor directionality (azimuth, elevation, roll). Sensor spatial separation for sensor pair. HRIR from MIT KEMAR or CIPIC subjects. Multiple sources, location specified as polar coordinates wrt. receiver. Order of reflections. Length of impulse response. Interpolation filter (used in sensor pair case). High-pass filter cut-off, for reduction of DC bias and low frequency ripple. Effect of distance (1/R) may be ON or OFF. Display Options Plot of surface absorption vs. frequency. 3D display of room, receiver and sources geometry. Plot of mean reverberation time (RT 60 ) vs. frequency. Plot of impulse response(s) vs. time or sample number, colour coded for left (L) and right (R). Magnitude spectrum corresponding to above impulse response(s), FFT length selectable, L/R colour coded and HRTF superimposed when MIT or CIPIC data selected. 2D zoom and rotatable plan view of room, receiver, source(s), surrounding image rooms, and image sources, with source intensity indicated by L/R colour code. 3D version of the above plan view with source intensity displayed as stem plot height, L/R colour coded. 3D zoom and rotatable view of room, receiver, source(s), surrounding image rooms, and image sources, with source intensity indicated by L/R colour code. Utilities for processing data files A convolution operation that accepts *.mat, *.au and *.wav format, and displays the impulse response data, audio file data and convolution result as line graphs, and additionally the latter two as spectrograms. An accumulation function for use in building cocktail party scenarios, also with signal displays. A converter between *.wav, *.au and *.mat, one and two channel file formats. A high-pass filter (4 th order Butterworth). A low-pass filter (4 th order Butterworth). File formats MATLAB *.mat, Microsoft Windows PCM *.wav, *.au 3

4 EDUCATIONAL USE OF ROOMSIM The Roomsim program may be of interest to educators delivering courses or modules in acoustic signal processing, music technology, auditory perception, psychoacoustics, room acoustics, and digital signal processing. It can form the basis for standard demonstrations, exploratory classroom laboratories or undergraduate projects. For example, the effect of surface and air absorption, room volume and air humidity can be demonstrated, and since the user has control of surface absorption, the image source method can be effectively illustrated by adding room surfaces one at a time and controlling the order of reflections displayed. The simulation may be used as a reverberation chamber for generating audio effects or demonstrating the complexity of audio signals from reverberant enclosures. The acoustic comb filter effect can be demonstrated by placing a sensor close to a reflective surface, and the spectral magnitude plot examined to confirm the notch frequencies predicted by calculation. The difference in sound quality and perceived sound stage localisation between recordings using a stereo microphone pair or a binaural head can be demonstrated. The 3D graphic displays of room geometry, receiver and primary source locations, image rooms and image sources (location and strength) are powerful visualisations of the complexity of the reverberant environment. The BSc Hons in Music Technology at the University of Paisley The Music Technology programme is currently delivered over four sessions as six core and two option modules per session of two, fifteen-week semesters. The programme aims to satisfy sections of the UK Quality Assurance Agency Music Benchmark by integrating theory and practice, and developing themes including, audio engineering, video techniques and multimedia. Supporting modules provide a sound theoretical underpinning of digital music, professional protocols and practices, detailed knowledge of the current technological base and an appreciation of future trends, insight into related fields and the context in which music technology can be applied to these. Presentation is through a range of teaching and learning methods which include: Lectures, personal research and independent learning; small group teaching, group project work and one to one interaction; peer learning through tutorial discussion of colleagues work; studio and computer laboratory work including hands-on experience of the use of electronic equipment for composition and recording; field work in engineering live sound; and external placements. An engineering ethos is developed in those modules that discuss equipment requirements, specification, design and implementation. The majority of students enter this programme either into first year (Level 7) from school or into third year (Level 9) from Higher National Diploma courses at Further Education Colleges and appropriate overseas institutions. Students are required to demonstrate an integrated and detailed understanding of technological, computational, theoretical, cultural and aesthetic systems, terminologies and conventions in the context of Music Technology. In particular they should acquire and demonstrate specialist knowledge and understanding of the underlying principles of digital audio and the use of computation to synthesise and process sound. More general educational aims are to develop critical, analytical problem-based learning skills and the ability to communicate clearly and concisely, i.e. the transferable skills to prepare the student for graduate employment undertaking creative activities within the technological sectors of the music and media industries. An example coursework The author uses the Roomsim program within the Advanced Audio Signal Processing module delivered in the second semester of the Honours degree year (Level 10) of the Music Technology programme. As with other modules on this programme contact time over the twelve teaching weeks is set as a block of four hours per week and in this module that is utilized as, on average, 1.5 hr of lecture and 2.5 hr of practical laboratory work. The laboratory work contains formative exercises intended to familiarize students with the Roomsim package and to build the experience and skills necessary for a major assessed coursework that is a required submission and comprises 30% of the module mark total. Coursework Specification (as issued to students) Using MATLAB, the Roomsim package, audio editor applications, and referring to any other appropriate resources, submit a CDROM containing Roomsim setup files, audio examples and analytical data from an investigation of the stipulated simulated recording and replay scenarios. The experimental element comprises two major parts: a simulation of three recording scenarios 4

5 in each of two recording locations, and a simulation of two replay scenarios. In all cases, the receiver will be placed on, and directed along, the centre line of the room (in plan view) parallel to the x-axis at the 1.2 m average ear height of a seated person. The recording locations are a sound treated studio room and a small hall. The studio has high absorbency walls, suspended acoustic tiled ceiling, and wooden platform floor. The hall is curtained down one wall, has two plastered walls, one gypsum boarded wall, suspended acoustic tiled ceiling and a wooden platform floor. It has a stage 0.7 m high at one end that adds to the performers heights but is not otherwise acoustically modelled. The performers are a seated instrumentalist and a standing vocalist. The instrumentalist can be considered as a sound source at an effective height of 0.7 m plus any stage height. The vocalist can be considered as a sound source at an effective height of 1.3 m plus any stage height. In both scenarios the vocalist is on the centre line of the space (in plan view) parallel to the x-axis and the instrumentalist is 1 m laterally in the +ve y direction from the vocalist. Three methods of recording are to be simulated at each location: (i) A pair of co-incident crossed dipole microphones (i.e. Intensity stereo). (ii) A NOS configured pair of cardioids (i.e. Intensity stereo plus time delay information). (iii) A Kemar dummy head as characterised by MIT (i.e. binaural stereo). N.B. Two recording venues, times three receiver types, requires six Excel setup files. Two methods of replay are to be simulated: (i) Headphone listening i.e. replay of the stereo files from each of the recording scenarios. (ii) Listening in a simulated domestic room via a pair of stereo loudspeakers using as the receiver one of the CIPIC dummy heads. The stereo files from the recording scenarios need split into two mono files to yield the audio material for the left and right loudspeaker sources. Coursework materials The coursework materials are provided via the university s Blackboard electronic teaching support system, and may be accessed within or outside a campus. The students are supplied with a partially completed Excel spreadsheet setup file for Roomsim which they must complete using a provided data sheet containing important parameters of the recording and replay scenarios. Flowcharts and a logical format for file naming are provided to guide their operations and ensure a coherent submission. The audio data provided are two unrelated mono recordings, one a solo vocalist and the other a solo instrumentalist. This material is chosen such that some significant gaps exist to aid acoustic location of the individual performers and that the composite performance has some musical validity. Each student is issued a different CIPIC head to enforce uniqueness of a major section of their coursework. In addition to performing the simulations and submitting all important data files, the students are required to complete two tables provided as part of the assessment [Table 2]. One requires them to extract quantitative data from the plots produced by Roomsim, the other requires them to form scored qualitative judgments on particular properties of their recordings as heard in the replay scenarios. The educational purpose of the tasks is to encourage the application of a reflective scientific/engineering approach to the critical judgment of audio materials based on the analysis of data. Student experience Students entering the Advanced Audio Signal Processing module have no prior experience with the Roomsim package. They have access to it in a computer laboratory and may download and install the compiled version on their laptop or home computers without requiring a MATLAB installation or license. Student feedback from two cohorts indicates that the combination of learning to operate a large software package and using it to perform a demanding exercise within an environment of standardized procedures and reporting, is judged relevant to their educational development and employability skills. In the first year of operation, the average mark achieved for this coursework over nine student submissions was 77%. A similar performance is anticipated from experience with the current similar sized cohort. 5

6 Table 2.- Quantitative and qualitative judgements required (extract) Reverberation (Impulse Response 10LOG10 Plot) Frequency Response (Reverberation Time vs. Frequency plot and Magnitude spectrum plot) RT60 Sabine (simple) [sec] Estimated IR length to - 60dB [sec] Ratio of Estimated IR length to Sabine RT60 Region(s) of major boost or cut in Reverberation Time vs. Frequency plot [Hz] Region(s) of major peaks in room freq response (>3dB) [Hz] Region(s) of major dips in room freq response (>-3dB) [Hz] Evidence of comb filtering? (If YES, identify notch & freq response plot(s) ) Qualitative Judgement (Score) Performer apparent separation (1 Obvious, 2 Not sure, 3 Definitely none) Stability of separation (1 Stable, 2 Unstable) Vocalist apparent distance from listener re. accompanist (1 Equal, 2 Closest, 3 Farthest) Relative dominance of vocalist (1 Well Balanced, 2 Dominant, 3 Dominated) Impact of venue reverberation (1 Pleasant, 2 Acceptable, 3 Problematic) CONCLUSION The Roomsim program is a free, open-source educational tool for educators and students in audio signal processing, music technology, auditory perception, psychoacoustics, room acoustics, and digital signal processing. Students can be set more or less demanding exercises by varying the amount of setup data provided and the depth of interpretation required. Roomsim also features sufficient flexibility to be a useful tool for researchers working in areas of audio signal processing who require to generate monaural/binaural audio signals that are sufficiently realistic for the assessment of signal processing algorithms aimed at improving the performance of e.g. speech recognisers, hands-free telephones and hearing aids. Acknowledgements: The author acknowledges the contributions of: Kalle J. Palomäki (Laboratory of Acoustics and Audio Signal Processing, Helsinki University of Technology, Finland), Guy J. Brown (Department of Computer Science, University of Sheffield, U.K.) and Derek Turner (School of Computing, University of Paisley, U.K.). The author also gratefully acknowledges permission to include in Roomsim the HRIR data from CIPIC at the University of California [8] and the MIT Media Laboratory [9]. References: [1] May [2] Allen, J. B. and Berkley, D. A., "Image method for efficiently simulating small-room acoustics", JASA 65(4), , [3] Palomäki, K. J., Brown, G. J., and Wang, D. L., A Binaural Processor for Missing Data Speech Recognition in the Presence of Noise and Small-Room Reverberation, Speech Communication 43, , [4] Peterson, P. M., "Simulating the response of multiple microphones to a single acoustic source in a reverberant room", JASA 80(5), , [5] Gardner, W. G., and Martin, K. D. HRTF measurements of a KEMAR dummy head microphone, In Standards in Computer Generated Music, Haus, G. and Pighi, I. eds., IEEE CS Tech. Com. on Computer Generated Music, [6] Algazi, V. R. et al., The CIPIC HRTF Database, IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, [7] May [8] May [9] May