Item

Abstract

Across musical genres worldwide, there are many styles where the
shortest conceptual units (e.g., notes) are often performed with
ornamentation rather than as static pitches. Carnatic music, a style
of art music from South India, is one example. In this style, ornamentation can include slides and wide oscillations that hardly
rest on the theoretical pitch implied by the svara (note) name. The
highly ornamented and oscillatory qualities of the style, in which
the same svara may be performed in several different ways, means
that transcription from audio to symbolic notation is a challenging
task. However, according to the grammar of the Carnatic style,
there are a limited number of ways that a svara may be realized in
a given raga ¯ (melodic framework), and these ways depend to some
extent on immediate melodic context and svara duration. Therefore,
in theory, it should be possible to identify not only svaras but also
the various characteristic ways that any given svara is performed -
referred to here as ‘svara-forms’.
In this paper we present a dataset of 1,530 manually created
svara annotations in a single performance of a composition in raga ¯
Bhairavi, performed by the senior Carnatic vocalist Sanjay Subrahmanyan. We train a recurrent neural network and sequence
classification model, DeepGRU, on the extracted pitch time series
of the predominant vocal melody corresponding to these annotations to learn an embedding that classifies svara label with 87.6%
test accuracy. We demonstrate how such embeddings can be used
to cluster svaras that have similar forms and hence elucidate the
distinct svara-forms that exist in this performance, whilst assisting in their automatic identification. Furthermore, we compare the
melodic features of our 54 svara-form clusters to illustrate their
unique character and demonstrate the dependency between these
cluster allocations and the immediate melodic context in which
these svaras are performed.