NonContact Surface Metrology for Preservation and Sound Recovery from Mechanical Sound Recordings

1

• Non-Contact Surface Metrology for Preservation
  and Sound Recovery from Mechanical Sound
  Recordings
• J.W.McBride P.Boltryk
• University of Southampton
• School of Engineering Sciences

2
The Project

• UK Engineering and Science Research Council
  Funded, over 4 years. 2005-09.
• J.W.McBride
• M.Hill
• Peter Boltryk (Researcher Hardware and Signal
  Processing)
• Anthony Nace (Ph.D on Signal Recovery Methods)
• Samantha Zhao (Ph.D on Optical Sensing)

3
Collaborators

• TaiCaan Technologies. Ltd.
• Lawrence Berkeley National Lab. USA
• Point Source Ltd.
• Will Prentice Nigel Bewley
• British Library.

4
The Sound Archive Project Web Site

• http//www.sesnet.soton.ac.uk/archivesound

5
Outline

• Introduction
• Current Systems and Sensors
• New System Developments
• The Process
• Preservation data
• Sound retrieval
• Examples of studies
• New Directions
• Conclusions

6
Introduction

• The aim is to create digitised surface maps of
  recorded surfaces which will then become the
  PRESERVATION copy.
• The methods will be applied to delicate, broken
  or valuable examples of media.
• It is likely that stylus methods will for the
  near future remain the main method for ACCESS

7
Systems IRENE (2D system)
8
IRENE
9
Sensors for 3D measurements

• The Aim of the project is to provide 3D surface
  maps of recorded surfaces which can be regarded
  as preservation copies, and act to preserve the
  original recording.
• The full surface should be represented as 3D
  data.
• Areal Measurements Interferometers
• Triangulation Systems
• Confocal Point Sensing systems
• Laser Focusing
• Chromatic Abberation

10
Interferometer Measurements
11
Point Sensing Systems
12
XY Surface of Edison Cylinder
13
2D Section of Data
14
Sensor study TL sensor principle

• Detects deflection of incident light on CCD array
• Highest sample rate for sensors considered in
  study 2kHz
• Large 30µm light spot this study highlights
  difficulty in resolving small features
• High point-to-point measurement noise

15
Sensor study CL sensor principle
Red laser source

• Collimator lens oscillated to vary focal point
  through sensors gauge range
• Lens position corresponding to highest received
  light intensity during measurement cycle related
  to surface height

Beam splitter
Tuning fork
time
A
B
C
16
Sensor study WL sensor principle
Polychromatic source

• Polychromatic light focused onto surface
• Lenss chromatic aberration focuses distribution
  of wavelengths at calibrated range of focal
  heights along axis of sensor
• Surface height calculated by detecting wavelength
  corresponding to spectrometers peak intensity

Beam splitter
Spectrometer
Lens
I
Surface
?2
?
17
Sensor study sensor comparison parameters
Table 1 Comparison of Critical sensor parameters
18
Sensor study (6) reference cylinder design
19
Sensor study (7) reference cylinder (diamond
cut)
Figure Cross-section of five 30µm deep
diamond-cut grooves measured using TL, CL, WL
sensors respectively, again compared with the
stylus profilometer Example regions labelled
(A) show WL sensors data loss on steepest
angled sides Note high level of noise affecting
TL sensor output
20
Data Format

• Each Z value is a 15bit representation of height
• 96kHz sampling for preservation data, requires
  0.01 increments for a 160 rpm cylinder. For a 2
  inch cylinder the spacing in 4-5µm.
• The scanning axis require a similar grid
  spacing. (10µm used here).
• Therefore a 90-100mm cylinder will be mapped with
  more than 300 Million data points.
• The data files are approx. 9GBytes, for a 4
  minute recording.
• This is reduced to less that 1GByte by data
  reduction.

21
Sensor study total scan time
Table 2 Comparison of time taken to scan Blue
Amberol cylinder
Note this is for benchmark sections only
actual cylinders recorded length is 92mm
22
Required resolution
Vocal component
Vocals vibrato
Instrumental
23
Dynamic Images of Spectrum
24
Required resolution (4)

• To resolve groove modulations of this magnitude,
  we require sensor resolutions of 20 30 nm
• This specification pushes us to the limits of
  current metrology technology
• Averaging is required for CL sensor increasing
  scan times

25
Measured groove profile - WL
26
Measured groove profile - CL
27
3D Solution for Disc Recordings

• The cylinder problem is compounded here with
  greater surface areas.
• Initial studies will be based on the commercial
  WL system with radial scanning
• Initial studies have shown that the greater angle
  of the groove profile will lead to difficulties
  with the process.

28
Future developments air bearing system

• To be used for preservation scans of flat disc
  media
• Air bearing system offers advantages of flatness
  of travel, speed (scan times) and lateral
  resolution

29
The Sound Generation Problem

• Many applications rely on accurate image
  segmentation of fine details
• Automatic detection of ridges and valleys common
  for
• Biomedical Imaging
• Forensics and biometrics (fingerprinting, facial
  recognition etc.)
• Geographical surveys / Hydrology

30
Sound Extraction Procedure

• 1. Data Import Produce large matrix of surface
  heights.

Z - Dimension of Groove Height
T - Dimension around Cylinder Circumference
X - Dimension along Cylinder Axis
31
Sound Extraction Procedure

• 1. Data Import This process produces unraveled
  cylinder

32
Sound Extraction Procedure

• 2. Interpolating missing data along linescan

Relative Height (mm)
Distance Along Linescan (mm)
33
Sound Extraction Procedure

• 3. Deduce Groove Spacing

Autocorrelation function used to identify
periodic structure.
x(n)
d
d
Typical Groove Structure
Autocorrelation of Groove Structure
34
Sound Extraction Procedure
4. Deduce All Local Groove Minimas and form
Feature map
Minimas found by examining lowest point in
neighbourhood or by examining the value of
gradient.
35
Sound Extraction Procedure

• 4. Deduce All Local Groove Minimas and form
  Feature map

Start
End
Feature Map is unravelled cylindrical surface
36
Sound Extraction Procedure

• 5. Determine Groove Starting Positions

37
Sound Extraction Procedure

• 6. Track Groove Paths

These trajectories are used to index regions on
the surface containing the audio
38
Sound Extraction Procedure

• 7. Create Coherent Time Series

z(t)
Direction along complete track
Groove displacement
39
Sound Extraction Procedure
8. Differentiate Groove Displacement

• Playback signal is approx. proportional to stylus
  velocity
• To simulate stylus motion, differentiate
  displacement with respect to time, to get
  velocity.
• If z(t) is groove displacement, then

Audio Waveform
40
Sound Extraction Procedure
9. Filtering And Equalisation

• The Mastering of sound is subjective.
• In general, wave files are band pass filtered
  300Hz 5kHz (accounts for bandwidth of
  recording system)
• De-clicking, De-popping, De-noising can
  then be applied if necessary.

41
Savitzky-Golay Smoothing
42
Minima detection
43
Minima detection
44
Minima detection
45
Minima detection
46
Implications for Audio Signal

• Boundaries at 0 and 360? do not match up
  exactly due to measurement process.
• CLICKS every rotation.

0?
360?
47
Implications for Audio Signal
Unwanted interpolation of debris (may cause
impulsive noise)
48
Stylus Transfer
49
Optical Recovery signal
50
Test cylinder SNR / THD comparison

• Test cylinder recorded with pure tones by
  electrical cutting stylus
• SNR comparison
• Scanned by optical method (grooves therefore
  unplayed by reproducing stylus)
• Transferred by stylus at British Library Sound
  Archive
• Compare audio signals achieved by each method for
  each tone frequency
• Cylinder was un-cleaned after cutting so dust was
  measured
• Groove profile rougher than mass-produced Blue
  Amberol examples

51
Aural and spectrogram comparison

• Audio file includes 4 signals,
• 250Hz tone recovered by optical transfer
• 250Hz tone recovered by stylus transfer
• 1.6kHz tone recovered by optical transfer
• 1.6kHz tone recovered by stylus transfer
• Signals are normalised with reference to each
  signals rms value
• Audio data declicked in software
• Band pass filtered in 150 20kHz using 4th order
  Butterworth filter
• Spectrogram derived from (3) and (4) 1.6kHz tones

52
(No Transcript)
53
SNR and THD metrics

• Based on VisualAudio Calculations Johnsen et al,
  JTS 2004
• Two metrics Signal-to-Noise Ratio (SNR) and
  Total Harmonic Distortion (THD)
• Signal-to-Noise Ratio
• Psignal is power in signal band of interest (f0
  5Hz),
• Pnoise is power in noise band, 200 Hz lt freq lt 20
  kHz, but not including signal band Psignal
• SNR calculated by
• Total Harmonic Distortion
• Pharmonics is power of signal in first 3
  harmonics, in 5Hz bands
• Psignal is power in signal band of interest (f0
  5Hz)
• THD calculated using

54
SNR as function of tone frequency
SNR roughly comparable across the frequency
range Stylus transfer typically achieves 2dB
improvement over optical transfer
55
THD as function of tone frequency
Total harmonic distortion lower for optical
transfer
56
Study Identifying Wear from Stylus Playback

• Artefact Brown Wax Cylinder c.1888.
• Reported to contain the voice of Queen Victoria.

57
Identifying Wear

Main Groove Structure
58
Identifying Wear
Relative Height (mm)
130 µm
59
Groove Shape Profile
60
Effect of Additional Stylus
Bottom of Groove Has been deepened. Information
lost.
61
Other Examples of Wear
62
Solution?

• Recovering sound from inside wear region proved
  unsatisfactory.
• The Virtual Stylus can be placed anywhere in
  the groove.
• Observe groove features and recover sound outside
  of the wear region.

63
Groove Features
64
Feature Map
65
Feature Map
Positive Gradient
Groove Shape
Virtual Stylus
Negative Gradient
66
Her Majesty spoke a few words?...

• Words not intelligible, but definite periods of
  speech are audible.

Audio extracted outside of wear region
Audio extracted inside wear region
N.b. Both files indentically band-pass filtered
400-1800Hz
67
Main Observations to date.

• We are recommending that rare and fragile
  cylinder recordings should not be played back
  using a stylus instrument.
• The sensor resolution for 3D surface mapping of
  cylinder recordings should be 10nm.

68
Conclusions

• 18 months into the project we have demonstrated
  the first ever?, full recording of a cylinder
  using optical scanning methods.
• We have identified that 10s nm resolution is
  required for the sensing systems.
• We have identified the best sensing technology
  for the cylinders.
• We have shown how to process damaged media,(Q.Vic)

69
New Directions

• 3D system for Flat Disks, December 2007.
• Approached by EMI to undertake a study on a
  broken cylinder.
• Seeking European Partners for a bid to the EU for
  research funding.
• Please contact me, jwm_at_soton.ac.uk