| IEEE AASP Challenge: Detection and Classification of Acoustic Scenes and Events 2016 |
| http://www.cs.tut.fi/sgn/arg/dcase2016/ |
|
|
| Task 2 - Synthetic audio sound event detection (SASED) |
| http://www.cs.tut.fi/sgn/arg/dcase2016/task-synthetic-sound-event-detection |
|
|
| Credits: Grégoire Lafay (IRCCYN, Ecole Cntrale de Nantes, France) |
|
|
| Contact: dcase-discussions@googlegroups.com |
|
|
| -- |
|
|
| This README file documents the training and development data of the Synthetic Audio Sound Event Detection (SASED) task. |
|
|
| The training dataset is contained in the folder dcase2016_task2_train/ |
| The development dataset is contained in the file dcase2016_task2_dev/ |
|
|
| -- |
|
|
| The training dataset is composed of mono recordings of isolated acoustic events typically found in an office environment. 11 classes are available: |
|
|
| * clearthroat (clearing throat) |
| * cough |
| * doorslam (slamming door) |
| * drawer |
| * keyboard (keyboard clicks) |
| * keysDrop (keys dropped on desk) |
| * knock (knocking on door) |
| * laughter |
| * pageturn (paper page turning) |
| * phone (vintage phone ringing) |
| * speech |
|
|
| Each class is represented by 20 recordings. Files are named according to the class name, i.e. classXXX.wav where XXX is a three-digit, unique and non-consecutive number. The audio format of the recordings is raw PCM, sampled at 44100 Hz, 16 bit (CD quality). All samples are normalized to -12 dBFS (true peak). |
|
|
| -- |
|
|
| The development dataset has 18 simulated acoustic scenes built using the samples of the training dataset. The simulation process is controlled by three parameters: |
|
|
| * ebr: Event to Background Ratio |
| * nec: Number of active Events per Class |
| * poly: boolean indicating if the scene is polyphonic (poly=1; events may overlap each others), or monophonic (poly=0; events are not allowed to overlap each others) |
|
|
| Thus, each file name (.txt or .wav) is as follows: 'dev_1_ebr_XX_nec_Y_poly_Z', where XX={-6,0,6}, Z={0,1}, and Y={1,2,3} if Z=0 or Y={3,4,5} if Z=1. |
|
|
| Thus, the Event to Background Ratio varies from -12 to 0 dB, the scene can be monophonic or polyphonic. In the first case, the number of event varies from 1 to 3 and in the second from 3 to 5. |
| Note: the sound sources for the training and development sets are the same. |
|
|
| -- |
|
|
| The folder 'annotation' in the development dataset contains the annotations for the aforementioned acoustic scenes as a plain ASCII file. Each .txt file contains information about the onset, offset, and event class for each event in the scene, separated by a tab, and is structured as follows: |
|
|
| onset1 offset1 EventID1 |
| onset2 offset2 EventID2 |
| ... |
|
|
| where 'onset' refers to the onset time in seconds, 'offset' refers to the offset time in seconds since the beginning of the scene, and EventID is a string corresponding to the event class as shown in the list above. |
|
|
| -- |
|
|
| Folder 'sound' in the development dataset contains the mono signals of the simulated acoustic scenes. The audio format of the simulated scenes is: raw PCM sampled at 44100 Hz, 16 bits (CD quality). |
|
|
|
|
