1, 2 not indexed yet

2a. short-time Fourier analysis and applying Hann window.
2b. 
0:00 getting magnitude nad hpase
1:30 hand-drawn squeezed Hann window; uncertainty principle
13:00 Dirichlet kernel as FT of a square pulse
40:00 when can two peaks be resolved using Hann window


3a.

0:00 What is a spectral envelope
10:00 how the cochlea works
12:30 resonant frequency of place along the cochlea
14:15 how a single cochlear "filter" acts
17:00 jnd in pitch
20:00 masking
22:23 pitch discrimination as function of frequency
24:00 difference tones
26:40 critical bands and harmonics
28:00 does ear use difference tones among higher harmonics to hear pitch?

3b.
0:00 bonk help window
0:50 loudness is amplitude ^ 0.3
1:00 how bonk works
4:20 bonk output in silly sones
5:00 bonk spew mode
10:00 make a patch with bonk.  Thresh 1000 1000 to shut it up
11:15 turn on spew mode and send to array
12:40 noise has more power in high critical bands
14:30 bonk args to get similar to critical bands
16:47 bonk -nfilters 1000 crashed...
17:00 rebuilding patch
bonk~ -nfilters 50 -halftones 3 -npts 512 -hop 256
24:00 putting a straight sinusoid into bonk~
24:30 similarits measure for sounds
25:00 searchvec-v1
33:00 normalizing two sets of bonk spectra so that one can index the other
24:00 speculation about pitch tracking with bonk~

4a.
2:00 cosine transform
5:00 historical Fourier question.  Cosine transform as symmetrized F transform
5:30 cosine transform of interest in heat flow on a bar
15:00 can you use cosine transform to look at sound?  (I think not).
24:00 making a triangle wave and a square wave using cosinesum
30:00 basis functions as dot products of vectors
32:00 projection onto subset of basis vectors
35:00 distance between two functions (L^2) and projection
40:00 in discrete form, L^2 is just Euclidean distance squared
42:00 projection minimizes distance

4b. 
(missing - try to regenerate)

5a.
0:00 singular value decomposition applied to principal component analysis
16:00 accumulating correlation matrix of a set of points
24:45 analyze-voice.pd: accumulate points of a "bonk spectrum" for PCA
34:00 julia script to read bonk accumulation to get basis vectors
46:00 back to Pd to show the basis vectors
58:00 random walks and PCI
1:01:00 the cosines, not quite, because anchored at left to zero
1:03:00 try to subtract mean from random walks

5b. 
0:00 fixing subtraction of mean from random walk for principal components
5:00 compared to principal components of two corpi: Ainger, Shatter, and speech
7:30 equal loudness contours and their implications for bark spectrum
10:00 don't compare unequally "important" dimensions with PCA
17:00 saving an equal-loudness contour to correct bark spectrum
24:00 Ainger corpus with and without equal-loudness correction
27:00 eigenvalues don't seem to depend much on equal-loudness normalization
30:00 linear prediction
43:00 example of LPC: resonance of coffee mug struck with a drumstick
1:04:00 sound of struck cup and its LPC error signal
1:10:00 LPC math

6a.
0:00 finishing LPC
10:00 factoring the polynomial
14:00 frequency response of a filter from pole plot
17:00 bandwidth tied to peak height in all-pole filters
23:00 LPC as spectral envelope predictor
23:30 f1, f2 bad functions of time (they can appear and disappear)
26:00 f1/f2/f3 might be better way to find vowels from voice recording
30:00 singers change their vowels to project better
38:00 bad sample rate behavior in LPC
41:00 pitch detection
48:00 sigmund demo on voice recording w/o amplitude control
49:00 dropouts; wrong pitch; octave errors
52:00 sigmund pitch and amplitude control
55:00 back to bark cepstrum; if you divide out total level quiet sounds jump
1:00:00 imaginary tetrahedral representation of pitch space
1:05:00 chroma and pitch height
1:11:00 shepard tone example to torture sigmund~

6b.
0:00 summarize pyramid and chroma/height helix together as pitch manifold
7:00 yin pitch detector
20:00 Terhardt approach to pitch detection
43:00 pitch spectrum from filterbank (q=17) versus FFT
45:00 actually picking peaks in FFT (versus using it as a filterbank)
46:00 reasons vibraphones are hard to deal with
51:00 maximum liklihood, model as sinusoid plus noise
57:00 multiple probability spaces
1:00 more about octave ambiguity - how big a peak at f/2 do you need?
1:04 terhardt pitch spectrum doesn't give probabliity of one pitch vs. another
1:08 flutes are worst things for octave errors
1:09 question about antescofo vs. discrete-pitch score following

7.
0:00 Krumhansl and Parncutt tables in parncutt paper
0:30 probe tone technique used by Krumhansl
2:30 parncutt bottom curve
3:00 digression into pitch tracking via subharmonics
4:00 minor chord could come from subharmonics
5:00 take a major chord and make subharmonics of each to get harmonicity chroma
11:00 difference in leading tone relevance in major scale between two graphs
12:00 difference between psychoacoustic and theoretical consonance
13:45 scale as two tetrachords
16:00 problem with triads in history of music theory
17:00 parncutt's main claim, tonality is from 1 triad
23:30 when is an interval consonant or dissonant?
29:30 break
45:30 Lerdahl's GTTM and Krumhansl's test - modeling tonal tension by K.
47:30 "green dog" and "big green dog" graphs (green has to be after bog)
48:00 "the big green dog" all branches go to the left.
49:00 the words have different levels (dog is at teh root of the graph)
4:30 "ran fast"  (the big green dog ran fast) - ran is th apex now.
51:00 each node of graph is exactly one thing in the sentence.
51:30 c-b-c melody. last c is root.  different graph (first c is higher level)
53:00 claire de la lune 
54:30 P. 2 pf paper (fig 2, Lerdahl's explanation. Join becomes arrow)
56:00 each node is a note.
58:00 branch right is rising tension, branch left is resolution.
1:09:00 - Parncutt's scale-belonging (back to Fig. 3) to evaluate chords
1:11:00 - distance traveled to get to a chord (Fig. 5) 
1:26 - Lerdahl new music structure (5th and minor second)
1:28 - both Lerdahl and Cope think music is linguistic
1:31 - Krumhansl experiment - play up to a point, ask if that's a good stop
1:33 - does lerdahl-distance predict stoppability?
1:34:00 in chromatic period, only 4 levels instead of 5 (no more triad)
1:36:00 result: tantalizingly close but have to fiddle with each new example
1:37:00 usefulness of lerdahl
1:37:00 NMF and attack detection next time, and try for musical time

8a.