We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

eess.AS

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Audio and Speech Processing

Title: Time-Frequency Audio Features for Speech-Music Classification

Abstract: Distinct striation patterns are observed in the spectrograms of speech and music. This motivated us to propose three novel time-frequency features for speech-music classification. These features are extracted in two stages. First, a preset number of prominent spectral peak locations are identified from the spectra of each frame. These important peak locations obtained from each frame are used to form Spectral peak sequences (SPS) for an audio interval. In second stage, these SPS are treated as time series data of frequency locations. The proposed features are extracted as periodicity, average frequency and statistical attributes of these spectral peak sequences. Speech-music categorization is performed by learning binary classifiers on these features. We have experimented with Gaussian mixture models, support vector machine and random forest classifiers. Our proposal is validated on four datasets and benchmarked against three baseline approaches. Experimental results establish the validity of our proposal.
Comments: 4 pages, 16 figures
Subjects: Audio and Speech Processing (eess.AS); Sound (cs.SD)
Cite as: arXiv:1811.01222 [eess.AS]
  (or arXiv:1811.01222v1 [eess.AS] for this version)

Submission history

From: Mrinmoy Bhattacharjee [view email]
[v1] Sat, 3 Nov 2018 14:07:41 GMT (2008kb)

Link back to: arXiv, form interface, contact.