Acoustic Scene Analysis

2018-10-07, Sunday

faaw icv

A new module has been implemented and added to FAAW.

The purpose of a new tool is to determine how attributes of selected representation vary between classes. Such functionality can be used to select attributes with small variability inside a selected class and with high variability between different classes.

2018-09-21, Friday

The challenge results are just published. Our system has been ranked at 39th position with accuracy equal to 65.8 % (64.3 - 67.4). The paper entitled "Audio Feature Space Analysis for Acoustic Scene Classification"  has been accepted and the results will be presented at the poster session of the DCASE workshop held in Woking, Surrey, UK.

2018-08-16, Thursday

ASA browser has been redesigned and rewritten. The new version called FAAW is still under development using Python/C++ languages.

Current features:

  • Highly configurable parametrisation stage. The acoustic features are computed in the C++ module, and at the moment, it supports 580 representations.
  • Convenient interface with mouse-free navigation.
  • The parameters of feature space are configurable in YAMLs files.
  • Editable text notes per audio file.
  • Easy access to audio files for inspection and annotation.
  • Various graphical representations of audio features.
faaw main window

2018-08-01, Wednesday

The results for Task 1A (classification of data from the same device as the available training data) of DCASE competition have been sent. The designed system uses traditional ML approach using carefully crafted audio feature space contained 223 attributes. Also, we have submitted a paper to the DCASE workshop.

2018-07-20, Friday

For this year DCASE Challenge, we have designed and implemented an auxiliary tool (ASA browser) to generate and analyse a feature space. The primary long-term goal is to support a semi-automatic process of finding associations between attributes and support the process of creating hierarchical features. The tool is in the early stage of development and is not available for download yet. At the current stage, the browser is capable of generating numerical values for the dataset and displaying several acoustical representations grouped by classes and locations.

The following screenshots show example representations for few audio files of the development dataset released for DCASE 2018 competition.

asa browser (1)
asa browser (2)
asa browser (3)
asa browser (4)
asa browser (5)
asa browser (6)