Introduction:

• Digital audio
– Recording
• Conversion: acoustic to electric
[Edison, Bell]
• Conversion: analog to digital
[Shannon, Bennett]
– Storage (or processing)
• Digital storage [CD, DVD, Blu-ray]
• Digital signal processing
– Playback
• Conversion: digital to analog
• Conversion: electric to acoustic
• ”Hi-res” audio
– Transparency
– DVDA, SADC, HD-Audio
 Introduction:
• The hearing
– 0 phon → treshold of audibility
– 120 phon → treshold of pain
– Bandwidth ~20kHz
• Transparency [ARA95]
– Dynamic range ≥ 120dB
• B ≥ 20 bits
– Bandwidth ≥ 26kHz
• fs ≥ 52kHz
• ”Hi-res” digital audio standards
– DVDA
• 24-bit, 192kHz
– SACD ISO226 Equal loudness curves
• 1-bit, 2.8MHz (120dB, 100kHz eff.)
– Intel HD Audio (Azalia)
• 32-bit, 192kHz
 Introduction:

• DAC
– Digital data → analog current or voltage
– Linear translation → LPCM
– For 24 bits: 223 ratio between MSB and LSB
• Or: 224 elements with thermometer encoding
• LPCM DAC impractical for audio
– Possible, but difficult and expensive
• Burr-Brown PCM 1702, 1704
• Can we reduce number of bits B?
DAC functionality
 Introduction:
• Requantization

Amplified 4 times

– Quantizer error → additive (independent) white noise source

• Additive noise approximation [Bennett]
• SNR: 6dB per bit
– Obviously not true, but allows for greatly simplified approximate analysis
 Introduction:
• Delta-Sigma Modulation
– Noise-shaped requantization
– Delta-Sigma Modulator (DSM) + oversampling = low in-band noise
– Additive noise model → simple analysis but hides idle-tone problem

Delta-Sigma Modulation Noise transfer function (NTF)

1bit mod1, 16xOSR

Amplified 20 times
NTF(z)=(z-1)
 Introduction:
• Delta-Sigma Modulation
– Generalized
– No delay-free loop → ntf[n]|n=0=0
• Up to seventh order filter seen in literature
• SACD → 5th order DSM and 1-bit quantization
– New problem: instability ( )
• Limits input range

1bit SACD compliant

Amplified 50.000 times!

 Introduction:
• This work:
– Exploring; focuses on issues of contention in previous research
– Pragmatic; assessing the relevance of previous research
• Motivation
– Build bridges between theoretical work and design practice
– Create a basis for future DSM activity at NTNU
– Work with audio!
• Results
– Five (six) papers published in or submitted to peer-reviewed journals and
conference proceedings
– Matlab-models
– Estimation methods
 Paper I:
• I.Løkken, A.Vinje, T.Sæther, ”Noise Power Modulation in Undithered
and Dithered High-Order Sigma Delta Modulators”.
– J.Audio Eng.Soc., vol.54, no.9, pp.841-854 September 2006
– Submitted Sept.2005, last revised June 2006
• Motivation
– Disputing claims about DSM dithering and noise power modulation (NPM) in
previous literature
– Simulation-based survey to assess NPM and practical implications (in-band)
 Paper I:
• What is dither and NPM?
– Applying noise at quantizer input makes quantization more random
– RPDF dither decorrelates quantization noise and signal
• Uncorrelated ≠ independent!
– 2*RPDF (TPDF) dither removes input dependency of error power
• Input dependent error power = noise power modulation
– N*RPDF dither removes input dependency of error statistical moments 1 to N
• Only moments 1 (mean) and 2 (power) seem audible. Since more dither means
more noise, TPDF is regarded as optimal for audio

RPDF dither TPDF dither

 Paper I:
• Dither in delta-sigma
– Popular claim: Delta-sigma modulator is self dithering
• Wrong
• Proved in Wannamaker thesis (and we heard it)
– Fact: DSM has same fundamental dither requirements as normal quantizer
• RPDF dither for no input correlation in error
• TPDF dither for no noise power modulation
– But: High-order DSM reduces harmful effects a great deal
• Negligible?
 Paper I:
• 1-bit DSM
mod1
 Paper I:
• 1-bit DSM
5th order SACD
 Paper I:
• Multi-bit DSM
3rd order 3-bit DSM
 Paper I:
• Multi-bit DSM
5th order 4-bit DSM
 Paper I:
• Multi-bit DSM
Relevant?
 Paper I:
• Results:
– This paper sought to provide pragmatic insight in the need for dithering
– 1-bit DSM → NPM inavoidable
• But in high-order modulator it is fairly benign
• Tonal behaviour is the dominant problem
• Dither to remove tones, but don’t over-dither
– Multi-bit DSM → NPM only removed with TPDF dither
• But in-band NPM is modest if the DSM is high order
• If quantization noise is made negligible, NPM is more of a theoretical than
practical concern
– Dithering may or may not be necessary
• Simulate!
• Keep in mind that dithering reduces analog SNR by reducing stable input range
 Paper II:
• I.Løkken, A.Vinje, T.Sæther, B.Hernes, ”Quantizer Nonoverload
Criteria in Sigma-Delta Modulators”
– IEEE Trans. Circuits and Systems Part II: Express Briefs, vol.53, no.12,
pp.1383-1387, December 2006
– Submitted June 2006, revised August 2006
• Motivation
– Non-overload method: Kenney, Carley, ”Design of multibit noise-shaping data
converters”, Analog Int. Circuits and Signal Processing, vol.3, pp.259-272
– P. Kiss, “Stable high-order delta–sigma DACs”, IEEE Trans. Circuits Syst. I,
Reg. Papers, vol. 51, no. 1, pp. 200–205, Jan. 2004.
• Suspicious claim: Non-overload only possible in error feedback DSM
• What is a non-overload (NOL) DSM?
– If the quantizer never overloads the DSM is guaranteed to be stable
– This is an extremely conservative requirement
• Most DSMs operate in overload
– NOL DSM needs quite many bits for good SQNR
• NOL DSM performance
– Needs quite many bits
– Not usable for low OSR

With RPDF dither With TPDF dither

Time que: 18m

 Paper II:
• Results:
– Shown that error feedback DSM is NOT inherently more stable than output
feedback DSM
– NOL analysis extended to include bounds for truncating and offset quantizers
– NOL requirements can be defined for all DSM types
• NOL DSM is guaranteed to be stable
• NOL DSM with appropriate dither only way to guarantee no tones or no noise
power modulation
– Non-overload requirement is very conservative
• High SQNR only for high OSR (>64) and high number of bits (>5)
 Paper III:
• I.Løkken, A.Vinje, T.Sæther, ”Segmented Dynamic Element Matching
using Delta-Sigma Modulation”
– Presented at the 31st AES International Conference, ”New Directions in High
Resolution Audio”, London UK, 25.-27.June 2007
– Submitted March 2007
• Motivation
– All multi-bit (B>1) DACs need dynamic element matching (DEM) to reduce
mismatch errors
– Many-bit DSM → Reduce DEM hardware complexity by segmentation
– Shape inter sub-DAC mismatch with dedicated Segmentation-DSM (SDSM)
• But: SDSM causes analog overhead
 Paper III:
• Segmented DEM
– Reduce complexity penalty by
segmenting DAC into sub DACs
– Problem: inter sub-DAC mismatch
• Modelled as leakage of truncation noise
– Solution: Dedicated SDSM
• Leakage of shaped noise instead
• But: Causes analog overhead
• Previous publications: error feedback
mod1 SDSM
– Least overhead
Aout  X  1     e
– Not very good shaping
– This paper explores higher order IIR
SDSMs using NOL method

Aout  X  1     eDSM
 Paper III:
• Segmented DEM.
– Simulations: Second-order SDSM
• DEM limited to second order
– Conservative second-order SDSM
shows clear improvement over mod1,
with much less overhead than mod2
 Paper III:
• Results/conclusions:
– Paper describes a useful utilization of NOL method
• Used to make hardware for segmented DEM with conservative NTF SDSM
– Paper presents an improved segmented DEM proposal
• 2nd order NOL SDSM performs much better than mod1-SDSM
• 2nd order NOL SDSM gives 30% less overhead than mod2-SDSM
– Segmented DEM allows for use of larger quantizers
• 8-10 bits instead of 3-6 bits typical today
• Most critical for low-OSR DSM, but also useful for hi-res as will be seen
– SDSM should have same order of shaping as DEM algorithm
• Current DEM algorithms → 1st or 2nd order
• No reason to ”over design” and create unneccessary overhead
• Error budgeting
 Paper IV:
• I.Løkken, A.Vinje, B.Hernes, T.Sæther, ”Error Estimation in Delta-
Sigma DA Converters”
– Submitted to Analog Integrated Circuits and Signal Processing, January 2008
• Motivation
– Difficult to find published estimates for how DAC errors are affected by the
DSM (NTF, OSR, number of bits…)
– In particular motivated by wanting to look at the jitter issue
• Results
– Estimation methods proposed for mismatch, jitter, switching errors
• All based on the additive noise model, i.e. not accurate or mathematically rigorous,
but simple and should give reasonable predictions
– Matlab models made to validate estimates with functional simulation results
• Instantiated through scripts using Schreier’s Delta-Sigma Toolbox™ to synthesize
modulators for changing parameter values
 Paper IV:
• What is jitter?
– Deviations in the sampling period
– Creates several kinds of distortion
• Sinusoid jitter → sideband distortion
• Random jitter → additional noise
– Jitter spectrum modulates (convolves) with signal spectrum
– Jitter sideband distortion not affected by DSM
– Jitter noise hugely affected by DSM

Error area model for jitter error simulation

Jitter error waveform

 Paper IV:
• How does jitter sound?

Jitter type Result

Original

1kHz sinusoid jitter

5Hz sinusoid jitter

White jitter

Exaggerated jitter audio examples

 Paper IV:
• Jitter estimate verification

50psRMS White jitter

50ps sinusoid jitter
 Paper IV:
• Other estimates

Signal-to-switching noise ratio (SSNR),

Signal-to-mismatch noise ratio (SMNR) 5ps switching asymmetry
 Paper IV:
• Results/conclusions:
– Paper describes simplified methods for DSM DAC error estimation
• X psRMS jitter, Y dB SNR spec → how many bits to we need?
• 5-bit this and that DSM, Y dB SNR spec → what jitter performance needed?
– Maximizing SQNR may have adverse effects on SJNR, SSNR
• Aggressive NTF makes the DAC more jitter susceptible
• Ditto for switching asymmetry
• Methods simplify trade-off
– Using many bits in the REQ is hugely advantageous for white jitter
susceptibility
• SJNR increases 6dB per bit
– Sinusoid jitter sideband distortion not affected by DSM parameters
• Justification for using many bits in DSM also for high-OSR audio
– Switching errors also improved with many bits
 Paper V:
• I.Løkken, A.Vinje, T.Sæther, ”Delta-Sigma DAC Topologies for
Improved Jitter Performance”
– Presented at AES 124th Convention, Amsterdam NL, May 2008 (poster)
• Motivation
– Many bits is beneficiary for jitter, but routing and DEM becomes complex
– Is using a semidigital filtering DAC (”FIR DAC”) a viable alternative?
– Estimation procedure developed in paper IV put into use for jitter
 Paper V:
• The semidigital DAC

M∙N levels out, compared with M∙N-level regular (segmented) DAC

 Paper V:
• Conclusions/results:
– Paper compares semidigital DAC with segmented DAC for many-level output
– Jitter immunity is best for the semidigital DAC, given reasonable mismatch
– Mismatch performance is better for segmented DAC (with 2nd order SDSM)
– Under the given set of design conditions the semidigital DAC is better
– Simple estimation methods extended to estimate performance in both cases
– Note: typo in eq.27, should say E{εmis[l]∙εmis[k]}=…
 Extra
• ”Paper VI”, not included since it was written after thesis handin
– ”Some Considerations for Spectral Analysis of Delta-Sigma Data Converters”
– Accepted for publication, ISAST Trans. Electronics and Signal Processing