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

Download:

Current browse context:

cs.SD

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Sound

Title: Multi-Speaker Localization Using Convolutional Neural Network Trained with Noise

Abstract: The problem of multi-speaker localization is formulated as a multi-class multi-label classification problem, which is solved using a convolutional neural network (CNN) based source localization method. Utilizing the common assumption of disjoint speaker activities, we propose a novel method to train the CNN using synthesized noise signals. The proposed localization method is evaluated for two speakers and compared to a well-known steered response power method.
Comments: Presented at Machine Learning for Audio Processing (ML4Audio) Workshop at NIPS 2017
Subjects: Sound (cs.SD); Audio and Speech Processing (eess.AS); Machine Learning (stat.ML)
Cite as: arXiv:1712.04276 [cs.SD]
  (or arXiv:1712.04276v1 [cs.SD] for this version)

Submission history

From: Soumitro Chakrabarty [view email]
[v1] Tue, 12 Dec 2017 13:17:30 GMT (752kb,D)

Link back to: arXiv, form interface, contact.