Page 1

:

Reconstruction of Mechanically Recorded Audio Signals using White-Light Interferometry

Presenter :Philippe GOURNAY

Authors : Khac Phuc Hung THAIPhilippe GOURNAYSerge CHARLEBOISRoch LEFEBVRE

Page 2

A joint project between two research laboratories at the Université de Sherbrooke:

GRPA - Speech and Audio Research Group (image and audio processing)

CRN2 – Nanofabrication and Nanocharacterization Research Center (measuring equipment)

Page 3

Reconstruct a digital audio signal from an Edison cylinder using non-contact profilometry

Earlier methods used a confocal probe which scans the recording medium one point at a time (Fadeyev 2005)

We use of a 3D optical profilometerwhich provides topographic information about an entire section in one operation

Page 4

OUTLINE

1. Blue Amberol Edison cylinders 2. White-light interferometry for surface

profiling 3. Audio reconstruction using 3D optical

profilometry 4. Experimental results

Page 5

1. Blue Amberol cylinders

Vertical cut recording process: audio signal = evolution of the depth of a continuous groove etched at the surface of the cylinder.

Spatial resolution needed to extract audio with Fe=48kHz:◦ Lateral 5µm

◦ Longitudinal 10µm

◦ Vertical 10nm

High resolution: micro-profilometry required

External diameter (mm) 55

Length (mm) 95

Rotation speed (rpm) 160

Groove density (tracks/cm) 79

Groove spacing (μm) 127

Groove depth (μm) 15-30

Audio duration Approx. 4 minutes

Audio bandwidth 150 Hz to 5 kHz

Page 6

2. White-light interferometry for surface profiling Principle

FOGALE Photomap 3D: 20x optical magnification chosen for our acquisition

Field of view: approximately 834 µm by 630 µm.

Resolution: at least 1µm laterally and 1nm vertically.

Page 7

3. Audio reconstruction using 3D optical profilometry

3.1 • Topographic data acquisition

3.2 • Surface curvature compensation

3.3 • Groove detection and segment extraction

3.4 • Segment concatenation

3.5 • Audio post-processing

Page 8

3. Audio reconstruction using 3D optical profilometry

3.1 • Topographic data acquisition

3.2 • Surface curvature compensation

3.3 • Groove detection and segment extraction

3.4 • Segment concatenation

3.5 • Audio post-processing

Page 9

3.1. Topographic data acquisition

Each measurement with the 3D profilometer: 1. A conventional (optical) grayscale microscopic image.2. A topographic image, also in grayscale but in which each

intensity corresponds to a certain depth in nm. Field of view : 615 µm by 520 µm; spatial resolution: 1 µm 5x5 median filter Down-sampled for 48 kHz

Audio sampling frequency 20% overlap between

measurements

Page 10

3. Audio reconstruction using 3D optical profilometry

3.1 • Topographic data acquisition

3.2 • Surface curvature compensation

3.3 • Groove detection and segment extraction

3.4 • Segment concatenation

3.5 • Audio post-processing

Page 11

3.2. Surface curvature compensation

Vertical lines in a topographic image appear at different elevations

Compensation using a simple application of Pythagorean theorem

Average “external” surface elevation before and after compensation:

Page 12

3. Audio reconstruction using 3D optical profilometry

3.1 • Topographic data acquisition

3.2 • Surface curvature compensation

3.3 • Groove detection and segment extraction

3.4 • Segment concatenation

3.5 • Audio post-processing

Page 13

3.3. Groove detection and segment extraction

Work on vertical lines in profilometric image Local minimum: elevation lower than 8 nearest neighbours Local minima must be at least 127 µm apart Use all vertical lines to build a short segment Some post-processing to reconstitute any missing or clearly

erroneous point.

Page 14

3. Audio reconstruction using 3D optical profilometry

3.1 • Topographic data acquisition

3.2 • Surface curvature compensation

3.3 • Groove detection and segment extraction

3.4 • Segment concatenation

3.5 • Audio post-processing

Page 15

3.4. Segment concatenation

20% overlap between adjacent measurement Correlate corresponding overlapping extremities to align

short segments Work on derivative signals because of a possible mismatch

in vertical alignment between consecutive measurements

Valley 1

Valley 2

Valley 3

x_offset

Page 16

3.4. Segment concatenation

Examples of correlation signals and corresponding overlapping extremities (note the vertical shift between images):

Page 17

3. Audio reconstruction using 3D optical profilometry

3.1 • Topographic data acquisition

3.2 • Surface curvature compensation

3.3 • Groove detection and segment extraction

3.4 • Segment concatenation

3.5 • Audio post-processing

Page 18

3.5. Audio post-processing

Simple 2-degree Butterworth high-pass filter: Fc = 800 Hz.

Simulate transfer function of a mechanical player + remove LF noise caused by surface deformations

Page 19

4. Experimental results

We scanned two turns from two different cylinders

Cylinders were manually turned and carefully repositioned between each measurement

One turn = approximately 350measurements

Allows us to reconstruct 3 complete groove turns, which correspond to 1.125 seconds of audio signal

+ used a vintage mechanical player (Amberola 30) to obtain reference signals

Page 20

4. Experimental results

Cylinder 1: Blue Amberol #2664 (My heart at thy sweet voice)

Reconstructed tonal signal

Page 21

4. Experimental results

Cylinder 2: Blue Amberol #3124 (With his hands in his pockets)

Reconstructed speech signal

Corresponding reference:

Page 22

4. Experimental results

Frequency content:

Temporalrepresentation:

Page 23

Conclusion

Good subjective quality of reconstructed audio signal

Acquisition still very long (40 seconds per measurement, 350 measurements for one cylinder turn = 1.125s of audio signal)

Limited audio bandwidth of the reconstructed signal (limited vertical resolution for complex surfaces?)

Thank you!