Comparative analysis of multiple musical performances

Paper by Craig Stuart Sapp2007 & 2008

Presented by Salehe Erfanian Ebadi

QMUL ELE021/ELED021/ELEM0215 March 2012

OutlineA technique for comparing numerous performances of an

identical selection of musicThe basic methodology is to split a one-dimensional

sequence into all possible sequential sub-sequences, perform some operation on these sequences, and then display a summary of the results as a two-dimensional plot

The current focus is on beat-level information for tempo and dynamics as well as commixtures of the two

MethodologyThe primary operation used on each sub-sequence is

correlation between a reference performance and analogous

The result is a useful navigational aid for coping with large numbers of performances of the same piece of music and for searching for possible influence between performances. segments of other performances

Collected over 2,500 recorded performances for 49 of Chopin’s mazurkas—on average over 50 performances for each mazurka

Some points about performancesKeeping track of differences and similarities between

performances is difficultA written score contains only the most basic of expressive

instructions (The unwritten rules of a composition are transmitted aurally between performers)

So to help exploring influences between performances tempo and dynamics are extracted from each, and then correlated against each other

Each performance in a plot is assigned a color

Raw DataBeat duration & Loudness (easier to extract) using Sonic

VisualizerMany other facets (note timings, voicing, articulation, …)

are ignored

Raw DataFor comparisons of musical dynamics between

performances, a smoothed version of the raw power calculated for the audio signal every 10 ms is sampled at each beat location

Loudness detection uses the smoothed data, with a delay of 70 ms because smoothing introduces a delay in data

Analysis ToolsNormalized (Pearson) Correlation

Values range from −1.0 to +1.0, with 1.0 being an exact match, and 0.0 indicating no predictable relation between the sequences being compared

Analysis ToolsScape PlotsCorrelation values are hard to interpret in isolationOriginally designed for timbral and harmony structural

analysis

Comparative performance scapesChoose one performance to be the reference for a

particular plotFor each cell in the scape plot, measure the correlation

between the reference performance and all other performances, then make note of the performance which yields the highest correlation value

Color the cell with a unique hue assigned to that highest-correlating performance

Timescapes

Timescapes

Dynascapes

Scape plots and parallel feature sequences

HYBRID NUMERIC/RANK SIMILARITY METRICS FOR MUSICAL PERFORMANCE

ANALYSIS

OutlineComparing numerical method for examining similarities

among tempo and loudness to Pearson CorrelationOther concepts such as “noise-floor” are used to

generate more refined measurements than the correlation alone

The measurements are evaluated and compared to plain correlation in their ability to identify performances of the same Chopin mazurka played by the same pianist out of a collection of recordings by various pianists

DataAlmost 3,000 recordings of Chopin mazurkas were

collected to analyze the stylistic evolution of piano playing over the past 100 years of recording history, which equates to about 60 performances of each mazurka

MethodologyLike before, beat timings extracted using Sonic VisualizerMarkup and manual corrections doneDynamics then extracted as smoothed loudness values

Other musical features ignored, yet important in characterizing a performance: pianists don't play right left hands together; legato and staccato hard to extract but important; tempo and dynamic, useful features (kept), allow listeners to focus their attention on specific areas

Derivations and DefinitionsType-0 score

Plain correlationType-1 score

Nearest neighbor performances in terms of correlation at all timescales

Type-2 scoreRemoving Hatto effect- removing best matches step by step

Type-3 scoreRemoving noise floor

Type-4 scoreOne additional refinement (taking the geometric mean)

Type-0 scoreThis type of correlation is related to dot-product

correlation used in Fourier analysisCorrelation values between extracted musical features

typically have a range between 0.20 and 0.97 for different performances of mazurkas

Though it's hard to only interpret similarity directly from correlation values

Type-0 scoreThe correlation values are consistent only in relation to a

particular composition, and these absolute values cannot be compared directly between different mazurkas

Different compositions will have different expected correlation distributions between performances

Type-1 score In order to compensate partially for this variability in correlation

distributions, scapeplots were developed which only display nearest-neighbor performances in terms of correlation at all timescales for a particular reference performance

Type-2 scoreScape displays are sensitive to Hatto effect:

If an identical performance to the reference, or query, performance is present in the target set of performances, then correlation values at all time resolutions will be close to the maximum value for the identical performance, and the comparative scape plot will show a solid color. All other performances would have an S1 score of approximately 0 regardless of how similar they might otherwise seem to the reference performance

To compensate for this problem, remove the best match from the scape plot in order to calculate the next best match

Type-2 scoreSchematic of nearest-neighbor matching method used in comparative timescapes

Type-3 scoreUsing the concept of noise-floor:

Definition of a performance noise-floor is somewhat arbitrary but splitting the performance database into two equal halves seems the most flexible rule to use

In any case, it is preferable that the noise floor does not appear to have any favored matches, and should consist of uniform small blotches at all timescales in the scape plot representing many different performers as is the example shown

Type-3 score

Type-4 scoreType-3 scores require one additional refinement in order

to be useful since performances are not necessarily evenly distributed in the feature space

Therefore, the geometric mean is used to mix the S3 score with the reverse-query score (S3r)

Type-4 score

EvaluationPresumably pianists will tend to play more like their previous

performances over time rather than like other pianists. If this is true, then better similarity metrics should match two performances by the same pianist more closely to each other than to other performances by different pianists

SourcesSapp, C. S. (2007). Comparative Analysis of Multiple

Musical Performances. In Proceedings of the 8th International Conference on Music Information Retrieval.

Sapp, C. S. (2008). Hybrid Numeric/Rank Similarity Metrics for Musical Performance Analysis. In Proceedings of the 9th International Conference on Music Information Retreival.

Thank You