by Scott Cannon Advisors: Dr. Jonathan Berger and Dr. Julius Smith Stanford Electrical Engineering 2002 Summer REU Program
Background The first phonograph was developed in 1877 as a result of Thomas Edison's research on two other inventions, the telegraph and the telephone. By 1908, the Edison Phonograph could play four minute recordings from 2 inch diameter wax cylinders which rotated at 160 rpm. The machines were handcranked, and used for music recording and business dictation purposes.
Background A collection of more than 1500 original cylinder recordings, and two fully functional phonographs were recently donated anonymously to Stanford University for use as an educational tool, and to be enjoyed by anyone via the internet. Photo: an early recording session, circa 1900
Goals The Edison Project focused on achieving two primary goals: 1) Digital encoding and preservation: To develop an efficient, high-fidelity means of capturing, processing, distributing, and preserving a collection of more than 1500 original Edison Cylinder recordings. 2) Audio Restoration: To research and design high-quality algorithms for removing distortion artifacts that are inherent to 90 year old recordings.
Part I: Method In Part I, research was focused to determine the best means of capturing and processing the audio from the phonograph. Every step of the signal chain was examined and constructed to maintain the highest fidelity. Focus of Part II Phonograph Microphone AD Conversion Computer Processing The Internet -The signal begins with an original handcranked Edison phonograph, nearly 100 years old. -In addition, our research also explored the use of professional cylinder players using today s technology. - Over a dozen microphones of various pickup patterns were auditioned in different locations around the phonograph. -All capture was done at 44.1 khz, and 16 bit (CD quality) -Selected recordings were captured using 24 bit encoding at 48 khz, for -A laptop was used for flexibility -Digital editing was completed with Wavelab, Matlab, SoundForge, and Protools software. Each recording received a mix of: -Noise reduction -Click removal -Equalization -Pitch Correction -Adjustments for eccentricity -Dynamic range compression Recordings were converted to.mp3 format for internet distribution, and cataloged in a detailed database at CCRMA.
Part I: Results A wide array of tests examined each component in the signal chain to create the best combination of equipment and tools. Frequency Response For example, it was discovered, that a stereo microphone configuration placed 1.5 meters from the phonograph captured 5 db less noise (around 1730 Hz) than a high-proximity microphone at 10 cm. Mic 10cm away Mic 1.5 meters away While this configuration helps to mitigate noise, the stereo recordings felt more distant to some listeners, who preferred a close microphone with acoustic dampening. Noisy Frequencies
Part I: Results Analysis showed that the phonograph produced no significant content above 4 khz. In addition, content below 80 Hz was negligible. Thus data compression algorithms and a bandpass filter were selected to deliver only the relevant content to the listener, resulting in a smaller file size for faster distribution over the web. Frequency Content of Captured Sound Frequency (Hz) Amplitude (db) Time (sec) The final system integrated the results of these and many other measurements at all points in the signal chain to capture the highest quality sound possible.
Part II: Method Part II of this research project aimed to develop a novel method to reduce the pops and clicks (very brief distortions caused by dirt and scratches) that are inherent to phonograph recordings. These signal distortions are eliminated through two primary processing steps: 1) Detection of the signal distortion. 2) Replacement of the distorted sound with an interpolated estimation.
Part II: Results Matlab was used to create algorithms that effectively detect substantial distortions, and splice good data from surrounding areas such that there are no signal discontinuities in the new waveform. The results sound very good! Amplitude (db) Time (ms) Original, distorted waveform Distortion detected and removed New, interpolated waveform Problem Area
Discussion While our algorithms for removing distortion performed well for distinct artifacts in the waveform, the ability to remove more subtle forms of audible distortion proved elusive with these methods. Future research will attempt to develop similar methods based on minimum variance estimation under various modeling assumptions to more effectively clean up old analog recordings.
Conclusion This research project accomplished two goals: Digital encoding and preservation: An efficient, high-fidelity means of capturing, processing, distributing, and preserving a collection of more than 500 original Edison cylinders was developed. Files are posted at http://ccrma-www.stanford.edu/groups/edison/ Audio Restoration: A method for removing signal distortion was created. A foundation has been built on which to develop more robust algorithms.
References Godsill, Simon J. and Rayner, Peter J.W. Digital Audio Restoration. 1998. Springer, London. Oppenheim, Alan V. and Willsky, Alan S. Signals and Systems, Second Edition. 1997. Prentice Hall. Lockhart, G. B. and Goodman, D. J. Reconstruction of missing speech packets by waveform substitution. In: Signal Processing 3: Theories and Applications. 1986. Garbit, Frederick J. The Phonograph and its Inventor, Thomas Alva Edison. 1994. Gunn, Boston.