1. Integrating Segmentation and Similarity in Melodic Analysis
Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Integrating Segmentation and Similarity in Melodic Analysis
Projects: Medieval Notation, Ear-Training Software
Tillman Weyde
ISE 599: Spring 2004

2. Integrated Segmentation and Similarity Model
Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Integrated Segmentation and Similarity Model
An analytical model for recognizing melodic structure
Integrates knowledge from music theory and empirical studies with system optimization by experimental data
Generates all possible structural “interpretations,” and rates them in order to select the most adequate one
The interpretations generated can be useful for music retrieval, music tutorials, and interactive music production tools
An extension of the system for rhythmic analysis
Computes and graphically displays results
“Experts” can add to interpretations with a graphical user interface
ISE 599: Spring 2004

3. Integrated Segmentation and Similarity Model
The recognition of melodic structure depends on both segmentation and similarity
Segmenting the melody into structural units (perceptual groups): Motifs [Motives]
Recognizing relations between motifs; determined by similarity
Segmentation and Similarity are inter-related, and a coherent computational model of melodic structure must integrate both aspects
Segmentation is influenced by the similarity relations of motifs in a melody
Similarity relations depend on how a melody is segmented

4. Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Motif [Motive]
A musical idea
either rhythmic, melodic, harmonic, or any combination of these three
may be as short as 2 notes, or long enough to consist of smaller units (also motifs, or cells)
has a distinct identity
A basic structural unit, which can be processed
can be sequenced, elaborated, or transformed (figuration)
often used in modulating passages to retain the melodic integrity
Classical development sections are typically built from motifs introduced earlier in the piece
Used to support or contribute to musical narratives (Leitmotif)
Motives can even create entire works.
ISE 599: Spring 2004

5. Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Motif [Motive]
3 motifs from Beethoven’s Pastoral Symphony
[no. 6, in F Major, op. 68, 1808]
Motives 2 and 3 can be said to be a reversal of rhythmic elements
ISE 599: Spring 2004

6. Interpretation Ratings
Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Interpretation Ratings
Interpretation Ratings are essential to the output of ISSM
The “quality” of each interpretation is determined by placing values on Segmentation and Similarity features
Segmentation features include: number of notes, duration of motifs, and pitch intervals at motif “boundaries”
Similarity features include: pitch, tempo, loudness, and contour
Segmentation Features correspond to the Gestalt Law of Proximity [events in close proximity are perceived as belonging to a unit]
Why is this 2 motives, and not 1?
ISE 599: Spring 2004

7. Interpretation Ratings
Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Interpretation Ratings
Segmentation of the melody
The ratios of average distance of the inner and outer intervals are calculated for each motif
For the outer notes, the minimal distance of interval notes in the circle of fifths is calculated
Fig. 3: Inner = 1 semitone average; Outer = 3.5 semitone average
ISE 599: Spring 2004

8. Interpretation Ratings
Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Interpretation Ratings
Assignment of related motifs based on Similarity
Global deviations and local deviations are rated separately
similar to Paradigmatic Analysis (Nattiez): the assignments represent how motifs are interpreted by listeners as being either identical or similar to preceding motifs
Interpretation of a simple melody
Global and Local pitch differences a motives ‘a’ and ‘b’
Jean-Jaques Nattiez, “Music and Discourse: Toward a Semiology of Music (1990). “All recurrent events belong to a Paradigm. Each unit is a “sign” held in relation to other signs without the connotation of meaning.
Semiology: Ferdinand de Saussure’s study of language [sign, signifyer, signified] There is no intrinsic relationship between signifyer (e.g. “chair”) and signified (concept of “chair”)
Musical semiotics can either be structuralist (Natteiz) or referential (Leonard Meyer)
Nattiez seperates music into 3 domains: Poietic (composer), Neutral (score), and Aesthetic (listener). He focuses on the Neutral.
ISE 599: Spring 2004

9. Interpretation Ratings
Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
Interpretation Ratings
Rating all possible interpretations is computationally inefficient because the possibilities grow exponentially with melody length, therefore:
A limited context of up to 10 notes is used
“Perceptually motivated constraints” are used to prevent implausible interpretations (Lerdahl and Jackendoff)
Calculation of the overall rating is done by a neural net defined by fuzzy rules [“neuro-fuzzy system”] and extended with a list processing features
Each connection of neurons corresponds to a fuzzy rule
Allows integration of prior knowledge with learning from data
Lerdahl and Jackendoff, “A Generative Theory of Tonal Music”: Theory of tonal music derived from generative linguistics. (Chomsky and Schenker) Surface and deep level structures.
“Grouping Structure = basic component of musical “understanding” expressed hierarchically (motive, phrase, section, etc.). This is perceived at a deep level--unlike metrical structure, which is perceived on the surface level.
ISE 599: Spring 2004

10. Learning from Data
ISSM learns from interpretation examples and uses these in an interactive training scheme
Interpretive training generates relative samples whenever the system chooses an interpretation that differs from one provided by an “expert”
The learning process changes the weights in the neural net

11. Weyde: Segmentation & Similarity in Melodic Analysis
Presentation by Brian Harlan
ISSM Modules
Rates all possible interpretations
Filters out according to constraints
Calculates the overall rating
Matches against adaptive system
“Training” generated relative samples
ISE 599: Spring 2004