Richard M Stern 18792 lecture August 28 2023 Department of Electrical and Computer Engineering Carnegie Mellon University Pittsburgh Pennsylvania 15213 Welcome to 18792 Advanced DSP Today will ID: 1031278
Download Presentation The PPT/PDF document "INTRODUCTION TO 18-792 ADVANCED DIGITAL ..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. INTRODUCTION TO 18-792ADVANCED DIGITAL SIGNAL PROCESSINGRichard M. Stern18-792 lectureAugust 28, 2023Department of Electrical and Computer EngineeringCarnegie Mellon UniversityPittsburgh, Pennsylvania 15213
2. Welcome to 18-792 Advanced DSP!Today willReview mechanics of courseReview course contentPreview material in 18-792 (Advanced DSP)
3. Important people (for this course at least)Instructor: Richard SternPH A26, (412) 916-7386, rms@cs.cmu.eduTeaching assistants: Ashwin Pillay, Oren Wrightapillay@andrew.cmu.edu, (412) 214-26534owright@sei.cmu.edu, (225) 276-0289Course management assistant: Cassandra PfannenstielHH 1113, 8-4951, cpfannen@andrew.cmu.edu
4. And, of course, the most important people are you!
5. Some course detailsMeeting time and place: Lectures here and now (Wednesdays as well)Recitations Friday 10:00 – 11:50, PH A18APre-requisites (you really need these!):Basic DSP course like 18-491/691Basic probability course like 36-217Some MATLAB or Python background (Stochastic processes not needed)Please chat with me after class if you have not taken 18-491 or 36-217 already
6. What topics in DSP do I really need to know?Relationships of DT representationsSample response/convolutionDiscrete-time Fourier transform (DTFT)Z-transform + ROCDifference equations + initial conditionsPole-zero locations + gain for one frequencyTopics related to the DFTDifference between the discrete Fourier transform and the DTFTLinear versus circular convolutionConvolving using the overlap-add and overlap-save methodsSignal flow diagrams
7. Does our work get graded?Yes!Grades based on:Machine problems and other homework (35-45%)Gradescope is now being used for all homework assignmentsMachine problems will be turned in using a standard formatThree exams (55-65%)Two midterms (October 11 and November 15), and final exam
8. Some details on homeworkIssued Thursday eveningsDue Friday at 0100 one week laterLowest homework grade droppedUp to 5 late days of late homework permitted, max 2 per weekStart your work over the weekend!
9. Other support sourcesOffice hours:2 hours per week for Stern, Pillay, and Wright, TBAYou can schedule additional times with me as neededCourse home page:http://www.ece.cmu.edu/~ece792 [not much there yet]Canvass to be used for grades (but probably not much else)Piazza to be used for discussions Faculty responses within 24 hours but not necessarily immediatelyGradescope to be used for homework assignmentsMATLAB/Python code will be turned in directly for execution
10. TextbooksMajor texts: Lim and Oppenheim: Advanced Topics in Signal Processing (out of print)Oppenheim and Schafer: Discrete-Time Signal Processing (from last semester)Newish: ADSP class notes!Material to be supplemented by papers and other sourcesMany other texts listed in syllabus
11. Diversity and inclusionEngineering is a classic profession for upwardly-mobile peopleHere and elsewhere we suffer from long-term endemic discrimination based on race, ethnic origin, religion, sexual preference, gender identity, disability, etc. etc.The multiple killings of black and brown people spring 2020 (and for a long time before and after) have sparked a good national discussion, although it is not clear if/when long-term change will happen.It is our collective responsibility to treat everyone fairly and equally, based on meritThe ECE Diversity, Inclusion, and Outreach committee’s home page ishttps://www.ece.cmu.edu/student-resources/dio.html
12. Register and Vote!!!!! Vote in Pennsylvania if you canAllegheny County voter registration page:https://www.alleghenycounty.us/elections/voter-registration.aspxRegistration deadline is Monday October 19 at 5 pmVoting by mail is possible
13. Academic stress and sources of helpThis is a hard courseTake good care of yourselfIf you are having trouble, seek helpTeaching staffCMU Counseling and Psychological Services (CaPS)We are here to help!
14. Academic integrity (i.e. cheating and plagiarism) CMU’s take on academic integrity:http://www.cmu.edu/academic-integrity/index.htmlECE’s take on academic integrity:http://www.ece.cmu.edu/programs-admissions/masters/academic-integrity.html Most important rule: Don’t cheat!But what do we mean by that?Discussing general strategies on homework with other students is OKSolving homework together is NOT OKAccessing material from previous years is NOT OK“Collaborating” on exams is REALLY REALLY NOT OK!
15. Advanced digital signal processing: some core conceptsThe biggest difference between 18-491/691 (DSP) and 18-792 is that ADSP is primarily concerned with processing random signalsKey application areas:Signal representationSignal modelingSignal enhancementSignal modificationSignal separation
16. Signal representation: why perform signal processing?A look at the time-domain waveform of “six”:It’s hard to infer much from the time-domain waveform
17. Signal representation: why perform signal processing?A speech waveform in time:“Welcome to DSP I”
18. A time-frequency representation of “welcome” is much more informativeImplementedIn Problem Set 4
19. Signal modeling: let’s consider the “uh” in “welcome:”
20. The raw spectrum
21. All-pole modeling: the LPC spectrum
22. Another type of modeling: the source-filter model of speechA useful model for representing the generation of speech sounds:PitchPulse train sourceNoise sourceVocal tract modelAmplitudep[n]
23. An application of LPC modeling: separating the vocal tract excitation and and filterOriginal speech:Speech with 75-Hz excitation:Speech with 150 Hz excitation:Speech with noise excitation:Comment: this is a major techniques used in speech codingImplementedIn Problem Set 9
24. Approach of Acero, Moreno, Raj, et al. (1990-1997)…Compensation achieved by estimating parameters of noise and filter and applying inverse operations“Clean” speechx[m]h[m]n[m]z[m]Linear filteringDegraded speechAdditive noiseClassical signal enhancement: compensation of speech for noise and filtering
25. “Classical” combined compensation improves accuracy in stationary environmentsThreshold shifts by ~7 dBAccuracy still poor for low SNRsCMN (baseline)Complete retrainingVTS (1997)CDCN (1990)–7 dB 13 dB Clean Original“Recovered”
26. Another type of signal enhancement: adaptive noise cancellationSpeech + noise enters primary channel, correlated noise enters reference channelAdaptive filter attempts to convert noise in secondary channel to best resemble noise in primary channel and subtractsPerformance degrades when speech leaks into reference channel and in reverberation
27. Simulation of noise cancellation for a PDA using two mics in “endfire” configurationSpeech in cafeteria noise, no noise cancellationSpeech with noise cancellationBut …. simulation assumed no reverb ImplementedIn Problem Set 10
28. Signal separation: speech is quite intelligible, even when presented only in fragmentsProcedure:Determine which time-frequency time-frequency components appear to be dominated by the desired signalReconstruct signal based on “good” components A Monaural example:Mixed signals - Separated signals - ImplementedIn Problem Set 4
29. Practical signal separation: Audio samples using selective reconstruction based on ITDRT60 (ms) 0 300No ProcDelay-sumZCAE-binZCAE-contImplementedIn Problem Set 5
30. Phase vocoding: changing time scale and pitchChanging the time scale:Original speechFaster by 4:3Slower by 1:2Transposing pitch:Original musicAfter phase vocodingTransposing up by a major thirdTransposing down by a major thirdComment: this is one of several techniques used to perform autotuningImplementedIn Problem Set 6
31. 18-792: major topic areasMulti-rate DSPShort-time Fourier analysisOverview of important properties of stochastic processesTraditional and modern spectral analysisLinear predictionAdaptive filteringAdaptive array processingAdditional topics and applications
32. Multi-rate DSPReview of sampling rate conversionPolyphase implementation of FIR filters for rate conversionMultistage implementations, with application to speech and music analysisDesign of quadrature and multi-channel filterbanks
33. Short-time Fourier analysisInterpretation as windowed Fourier transform or filter bankFilter design techniquesAnalysis-synthesis systemsApplications to speech and music analysisPhase vocodingManipulation of time and frequencyGeneralized time-frequency representationsWigner distributions and wavelet functions
34. Introduction to random processesStochastic process definitions and propertiesEnsemble and time averagesPower spectral density functions and their computationRandom processes and linear filtersGaussian and other special random processes
35. Traditional and modern spectral analysisIntroduction to statistical estimation and estimatorsEstimates of autocorrelation functionsTraditional approaches based on the periodogramPerformance of smoothed spectral estimatesNonlinear estimation: the maximum entropy methodParametric approaches to spectral estimation; linear prediction
36. Linear predictionLinear prediction using covariance and autocorrelation approachesLevinson-Durbin recursion and Cholesky decompositionDesign and interpretation of lattice filtersApplications to speech, bioinformation processing, and geophysics
37. Adaptive filteringIntroduction to adaptive signal processingObjective measures of goodnessLeast squares derivationsSteepest descentThe LMS and RLS algorithmsAdaptive lattice filtersKalman filtersMulti-sensor adaptive array processing and beamforming
38. Some possible additional topicsHomomorphic signal processing and the complex cepstrumBlind source separationSignal processing for speech analysis, synthesis, and recognition
39. Comment … one of my consulting cases in 2015(Andrea v Dell et al.)US patent 6,049,607
40. Comment … one of my consulting cases in 2015(Andrea v Dell et al.)US patent 6,049,607
41. SummaryLots of interesting topics that extend core material from DSPGreater emphasis on implementation and applicationsGreater emphasis on statistically-optimal signal processingI hope that you have as much fun with this material as I have had!
42.