Acoustical Society of America
158th Meeting Lay Language Papers
[ Lay Language Paper Index | Press Room ]
Time-frequency Test Signal Synthesis for Acoustical Measurements During Music Concerts
Ren Gang - garen@ece.rochester.edu
Dept. of Electrical and Computer Engineering,
Edmund A. Hajim School of Engineering and Applied Sciences,
University of Rochester
Rochester, NY, 14627
Mark F. Bocko- bocko@ece.rochester.edu
Dept. of Electrical and Computer Engineering,
Edmund A. Hajim School of Engineering and Applied Sciences,
University of Rochester
Rochester, NY, 14627
Dept. of Music Theory,
Eastman School of Music,
University of Rochester
Rochester, NY, 14604
Dave Headlam- dheadlam@esm.rochester.edu
Dept. of Electrical and Computer Engineering,
Edmund A. Hajim School of Engineering and Applied Sciences,
University of Rochester
Rochester, NY, 14627
Dept. of Music Theory,
Eastman School of Music,
University of Rochester
Rochester, NY, 14604
Popular version of paper 4pAA8
Presented Thursday afternoon, October 29, 2009
158th ASA Meeting, San Antonio, TX
We describe a new method for conducting acoustical measurements in concert halls during on-going musical performances. This addresses the long-standing problem in architectural acoustics of enabling the characterization of a performance space with an audience present, which has a significant effect on the acoustical properties of an auditorium. We employ adaptively designed test signals that depend upon the sonic field of the ongoing performance. Such test signals are broadcast and received in the music hall during the performance and the room response is inferred as in conventional acoustics measurements, however our test signals are hidden from the audience by the sound of the ongoing concert by employing auditory masking, the same psychoacoustic phenomenon employed in MP3 music file compression. By confining the test signal to time-frequency locations that lie beneath the masking threshold and then employing appropriately chosen filters, we can separate the response of the space to the test signals and infer room acoustical parameters through conventional techniques.
In conventional architectural acoustic measurements loud test signals are employed to excite a space and the resulting response is measured. Such measurements are an essential tool in "fine tuning" a new music hall and in studying existing fine music halls to better understand the properties that make them desirable. A significant problem in such studies is that the majority of measurements are conducted in empty music halls since it would be an imposition on an audience to conduct the tests in a filled venue. What is needed is a method to compete such measurements in an occupied performance space without imposing on the patrons.
The method we propose enables such a measurement. The sound of a musical performance has a complex time-frequency signature called a spectrogram. Our method takes advantage of a psychoacoustic phenomenon called auditory masking to place our test signals in regions of the spectrogram inaudible to the humans in the concert hall; however, these signals will be measureable using standard signal processing techniques. Figure 1 represents a spectrogram of a short time segment of a musical sound. The musical sound appears as the red peaks, which correspond to the fundamental component of a musical tone and its first two harmonic overtones. The tone is localized to certain frequencies but they extend over time, the third axis. The heights of the peaks indicate the energy level of the musical sound in the corresponding time-frequency locations. The energy distribution of musical sound in this time-frequency representation is sparse with significant open spaces in which to place our acoustic test signals (for example, the green bar) in the diagram. If the two signals (the acoustic test signal and the sound of concert) do not overlap in time and frequency it is fairly straightforward to separate them using standard methods of signal processing.
Figure 1. Masking of test signal for performing acoustic measurements during an on-going concert. Red peaks: the harmonic component of the sound of the concert. Blue surface: the masking surface, sound level below it is inaudible. Green bar: the acoustic test signal "hidden" under the masking surface. Yellow surface: the noise floor.
However there is a second requirement that the acoustic test signals should not be perceptible to the audience. To meet this requirement we take advantage of the phenomenon of auditory masking, in which the strong components of a sound can "mask" the weaker components. Figure 1 illustrates the masking threshold (blue surface) of the music (red), sounds below the masking surface are inaudible to a human listener. In general the masking threshold is highest near strong tones and is lower in regions where there is little spectral energy. Thus, if we limit our test signals to regions in the spectrogram beneath the masking surface the test signals can safely be concealed from the listener. In practice the energy level of test signals also must be above the noise floor (yellow surface) of the spectrogram so they are detectable by the acoustic test equipment.
Acoustics engineers can easily control the time-frequency energy distribution of their test signals to lie below the masking threshold and above the noise floor, however there remains another problem. The test-signal design process must take place in real-time throughout the concert, which requires that the engineer somehow must be able to predict the musical signal as the concert progresses. Specifically we need to predict the locations and intensities of the strong features in the spectrogram of the music. Fortunately in music, we do have the some ability to predict the future. One such method is based on a probabilistic formulation called a Markov random field (MRF), which basically states that the "future" of the on-going music is linked to its "present" and "history" by probabilities. So if we detect the beginning (onset) of a harmonic partial at some frequency location, its possible duration and future intensity can be predicted by analyzing the past behavior of the harmonic partials in the piece. These predictions are aided by having some knowledge of the compositional style of the music. The majority of musical compositions played in concerts belong to the category of tonal music, which is a highly predictable structured sound. However as we know from weather forecasts the reliability of predictions decreases the longer we attempt to look ahead, it is the same in this case. However to design test signals that may be masked by the music we need to predict only a second or two into the future. When the predictions begin to stray the measurements must be suspended for two reasons: first, the test signal might not be masked and would become audible to the audience; second, if the test signal or its reverberant response overlaps the on-going concert sound too much in the time-frequency spectrogram, the response would be difficult to separate. However during a concert that may extend for an hour or longer there are many opportunities to complete the measurements needed.