Image-Processing Methods Reconstruct Archival Audio Recordings
Investigators listen to music never meant to be heard again.
Point-and-click mp3 files make it hard to appreciate the first attempts to capture sound on glass, wax, tinfoil or even paper. Many of those records, albeit fragile and often noisy, survive. Carl Haber and his colleagues at Lawrence Berkeley National Laboratory in California, who developed image analysis tools for high-energy particle detection, have applied those methods to the task of maintaining and preserving sound recordings of historical importance.
Mechanical recording impresses the motion of a stylus in a medium. For traditional records, the stylus moves laterally as the media revolves, leaving a 2-D representation of the sound that can be imaged with standard microphotography. For such media as the original spools created by Thomas Edison, the stylus deflects vertically as the cylinder revolves, requiring a 3-D imaging method. Both 2- and 3-D reconstruction methods acquire a series of high-resolution images, connect the separate images, clean the combined image and then simulate the stylus response. The challenge for both methods is that information encoded in submicron structures covers hundreds of linear meters.
The accuracy required to reconstruct 3-D recording media, such as the Edison cylinder shown here, demands highly sensitive depth detection. Acquiring and stitching together thousands of confocal microscope images allows fragile media such as this to be read without any mechanical contact. Images courtesy of Lawrence Berkeley National Laboratory.
The resolution must preserve the audio bandwidth of the recording and detect amplitude variations down to the original noise level. After initial studies with a rectangular-format CCD imager, Haber’s team now uses a 4096-pixel linear CCD for 2-D line scanning. The optics image a roughly 2-mm length perpendicular to the record grooves; then the disc is rotated while the camera acquires 80,000 to 240,000 lines per revolution. The flat areas between the grooves and the bottom of the grooves retroreflect the coaxial illumination from a xenon lamp. Each groove is located by the four light-dark transitions within each cross section. At a typical scan speed of about 50°/s, a 10-in. shellac disc is scanned in 15 to 20 min. The image-processing routine connects the tracks, corrects for geometrical errors and removes noise.
The same image processing applies to 3-D data from Edison cylinders, but the images are acquired with a color-coded confocal microscope from Sciences et Techniques Industrielles de la Lumière (Stil SA) of Provence, France. As in standard confocal microscopy, incident light is sent through a pinhole conjugate to a given image plane and then collected through a spatial filter. By spreading the incident white light, the wavelength of the collected light directly indicates the distance. For a 127-μm-wide groove on a typical cylinder, the 7.5-μm spot will produce sufficient amplitude resolution scanned in 10-μm steps across the groove; rotating the cylinder in 1/100° steps gives an audio sampling rate of 96 kHz.
Working closely with the Library of Congress, the Berkeley team is evaluating the IRENE (Image, Reconstruct, Erase Noise, Etc.) disc-reading system for reconstructing and making accessible some of the Library of Congress’ 2.5 million recordings. Haber’s processing routines, applied to images from a commercial scanner, recently made news by reproducing sound from a phonautograph recorded on paper by the French inventor Édouard-Léon Scott de Martinville. The 1860 recording of Au Claire de la Lune is the earliest example of recorded sound, predating Edison by 17 years. Scott did not intend the music to be played back, only studied in its scribed form.
Image-processing methods adapted from high-energy particle experiments connect 80,000 or more individual line scans of damaged records to be combined, electronically cleaned and audibly reproduced.
Haber is developing a 3-D version of IRENE using the color-coded confocal microscope, which has improved to measure 180 points at 400 to 1800 Hz — nearly 200 times faster.
He enjoys applying techniques of optics, data processing, statistics and metrology to items of cultural significance. “This is a good example of how the methods of the physical sciences and engineering can be applied to a problem which, to some extent, is of interest to other fields of research.”
- A noncrystalline, inorganic mixture of various metallic oxides fused by heating with glassifiers such as silica, or boric or phosphoric oxides. Common window or bottle glass is a mixture of soda, lime and sand, melted and cast, rolled or blown to shape. Most glasses are transparent in the visible spectrum and up to about 2.5 µm in the infrared, but some are opaque such as natural obsidian; these are, nevertheless, useful as mirror blanks. Traces of some elements such as cobalt, copper and...
MORE FROM PHOTONICS MEDIA