CS 591 S1 Computational Audio

Size: px

Start display at page:

Download "CS 591 S1 Computational Audio"

Nathaniel Pitts
5 years ago
Views:

4/29/7 CS 59 S Computational Audio Wayne Snyder Computer Science Department Boston University Today: Comparing Musical Signals: Cross- and Autocorrelations of Spectral Data for Structure Analysis

1 4/29/7 CS 59 S Computational Audio Wayne Snyder Computer Science Department Boston University Today: Comparing Musical Signals: Cross- and Autocorrelations of Spectral Data for Structure Analysis Segmentation Computer Science Identifying structure from audio Arthur G. Lintgen: able to identify unlabeled recorded orchestral works by observing the spacing and patterns of grooves in an LP Inspired J. Foote (ISMIR, 2) to develop a MIR system based on structural similarity

4/29/7 Musical Form Units can be assigned letters (A, B, C) or functional names (intro, verse, chorus, bridge, etc) Strophic: repeats the same section, e.g. AA Binary: alternates two sections, which are often repeated, e.

2 4/29/7 Musical Form Units can be assigned letters (A, B, C) or functional names (intro, verse, chorus, bridge, etc) Strophic: repeats the same section, e.g. AA Binary: alternates two sections, which are often repeated, e.g. ABAB or AABB Ternary: third section is often a variation of the first, e.g. AABA, AABA, AA BA Arch: symmetric, repetition of sections around a center, e.g. ABCBA Rondo: main theme is alternated with sub-themes, e.g. ABACADA.. Variations: theme plus variations, e.g. AA i A ii AA iii Sonata: complex developmental form including the exposition, development and recapitulation of a given theme(s). Repetition Musical form is often defined by the amount of repetition across sectional units. Repetition is central to music (in harmony, melody, rhythm, etc). Significant variations are often found between repeated parts.

3 4/29/7 Repetition The information necessary to characterize repetitions is encoded in the feature vectors (e.g., chroma, spectrum, etc.) (a) A A 2 B B2 A 3 T C C 2 A4 A 5 (b) G# G F#.8 F E.6 D# D.4 C# C B.2 A# A A A 2 B B2 A 3 C 4 A 5 C 2 Audio Structure Analysis Given: CD recording Goal: Automatic extraction of the repetitive structure (or of the musical form) Example: Brahms Hungarian Dance No. 5 (Ormandy) 5 5 2

4 4/29/7 Basic Procedure Extract audio feature vectors (e.g., chromagraph) Cost measure and cost matrix self-similarity matrix Path extraction (pairwise similarity of segments) Global structure (clustering, grouping) Self-similarity matrix For an N-long sequence of feature vectors Self-similarity matrix: N x N matrix of pairwise (dis)similarities between vectors

Main diagonal: closer/ most similar values similar subsequences

5 4/29/7 Self-similarity matrix Vertical and horizontal axes represent time symmetric similarity function = symmetric matrix Main diagonal: closer/ most similar values similar subsequences (repetitions) -> diagonal stripes in the plot Self-similarity matrix

6 4/29/7 Self-similarity matrix Self-similarity matrix

7 4/29/7 Let s look at a similarity matrix and hear the piece of music to see how it represents the structure... Basic Procedure Self-similarity matrix Similarity structure 5 5 2

8 4/29/7 Basic Procedure Self-similarity matrix Similarity structure Basic Procedure Self-similarity matrix Similarity structure 5 5 2

9 4/29/7 Basic Procedure Self-similarity matrix Similarity structure Basic Procedure Self-similarity matrix Similarity structure 5 5 2

4/29/7 Basic Procedure Self-similarity matrix Similarity structure 5 5 2

10 4/29/7 Basic Procedure Self-similarity matrix Similarity structure Higher level structure can be determined by first extracting paths in the matrix

11 4/29/7 Path Extraction Thresholded, upper left Path Extraction Path removal

12 4/29/7 Path Extraction Path removal Path Extraction Extracted paths

13 4/29/7 Path Extraction Extracted paths after postprocessing Global Structure can be derived from relating paths at a higher level...

14 4/29/7 Music Synchronization: Lyrics-Audio Difficult task! Music Synchronization: Lyrics-Audio Lyrics-Audio Lyrics-MIDI + MIDI-Audio

4/29/7 System: SyncPlayer/LyricsSeeker High-Resolution Music Synchronization Normalized chroma features robust to changes in instrumentation and

15 4/29/7 System: SyncPlayer/LyricsSeeker High-Resolution Music Synchronization Normalized chroma features robust to changes in instrumentation and dynamics robust synchronization of reasonable overall quality Drawback: low temporal alignment accuracy Idea: Integration of note onset information

16 4/29/7 High-Resolution Music Synchronization Example: Beethoven s Fifth Audio MIDI Cost matrix Audio Warping path of poor local quality MIDI High-Resolution Music Synchronization Cost matrix windows are based on based on onset intervals, not uniformly spaced! Audio MIDI Cost matrix Audio Warping path based on onset information MIDI

17 4/29/7 High-Resolution Music Synchronization Ideas: Build up cost matrix with corridors of low cost Warping path tends to run through corridors of low cost note onset positions are likely to be aligned High-Resolution Music Synchronization Impulses zoom Decaying impulses zoom

18 4/29/7 High-Resolution Music Synchronization Cost matrix for decaying impulses High-Resolution Music Synchronization Cost matrix for decaying impulses Corridor of low cost

19 4/29/7 Music Segmentation Analysis Music segmentation pitch content (e.g., melody, harmony) music texture (e.g., timbre, instruments) rhythm How to find the musical sections of the piece? Music Segmentation Analysis Basic idea (from image processing) uses a kernel or mask to modify data points according to their neighbors Each data point is replaced by the weighted sum of its neighbors * kernel values This is given the convoluted name convolution in the signal processing world, with an excessively complicated definition. It is really just the same as a weighted sum of neighboring values, as with smoothing..

20 4/29/7 Music Segmentation Analysis Music Segmentation Analysis A binary kernel finds boundaries along the axis where sections of music different from neighboring sections

21 4/29/7 Music Segmentation Analysis The kernel is slid along the axis of the similarity matrix, and the value of the convolution is recorded for each time (in the center of the kernel): Music Segmentation Analysis The kernel is slid along the axis of the similarity matrix, and the value of the convolution is recorded for each time (in the center of the kernel):

22 4/29/7 Music Segmentation Analysis The kernel is slid along the axis of the similarity matrix, and the value of the convolution is recorded for each time (in the center of the kernel): Music Segmentation Analysis As the kernel is slid along the axis, the values calculated give us a novelty score for how much the music is changing at that point Different kernel types and sizes give a different perspective on the scale of the changes, from individual notes to large sections. Peak picking gives us the times where there is a potential start of a new segment of music:

23 4/29/7 Music Segmentation Analysis A Gaussian Kernel emphasizes changes at the center, and deemphasizes the edges (c.f. Hann Windows)

Audio Structure Analysis

Audio Structure Analysis Advanced Course Computer Science Music Processing Summer Term 2009 Meinard Müller Saarland University and MPI Informatik meinard@mpi-inf.mpg.de Music Structure Analysis Music segmentation pitch content