Sound Data Management Training » History » Version 4

Steve Welburn, 2012-07-04 05:47 PM

1 1 Steve Welburn
h1. WP1 2 Online Training Material
2 1 Steve Welburn
3 1 Steve Welburn
h2. C4DM Researcher use cases
4 1 Steve Welburn
5 1 Steve Welburn
h3. Quantitative testing - machine testing
6 1 Steve Welburn
7 1 Steve Welburn
bq. A common use-case in C4DM research is to run a newly-developed analysis algorithm on a set of audio examples and evaluate the algorithm by comparing its output with that of a human annotator. Results are then compared with published results using the same input data to determine whether the newly proposed approach makes any improvement on the state of the art.
8 1 Steve Welburn
9 1 Steve Welburn
Data involved includes:
10 1 Steve Welburn
* Software for the algorithm (which can be hosted on "Sound Software":http://www.soundsoftware.ac.uk)
11 1 Steve Welburn
* An annotated dataset against which the algorithm can be tested
12 1 Steve Welburn
* Results of applying the new algorithm and competing algorithms to the dataset
13 1 Steve Welburn
* Documentation of the testing methodology
14 1 Steve Welburn
15 1 Steve Welburn
Note that *if* other algorithms have published results using the same dataset and methodology, then results should be directly comparable between the published results and the results for the new algorithm. In this case, the methodology is already documented.
16 1 Steve Welburn
17 1 Steve Welburn
Also, if the testing is scripted, then the code used would be sufficient documentation during the research - readable documentation only being at publication.
18 1 Steve Welburn
19 1 Steve Welburn
If no suitable annotated dataset already exists, a new dataset may be created including:
20 3 Steve Welburn
* Selection of underlying (audio) data (which may be [[Copyright|copyright]] material)
21 1 Steve Welburn
* Creation of ground-truth annotations for the audio and the type of algorithm (e.g. chord sequences for chord estimation, onset times for onset detection)
22 1 Steve Welburn
23 1 Steve Welburn
h3. Qualitative testing - Listening tests
24 1 Steve Welburn
25 2 Steve Welburn
An example would be testing audio at various levels of compression using both standard techniques and a newly derived algorithm.
26 1 Steve Welburn
27 4 Steve Welburn
e.g. MUSHRA type tests ("Wikipedia":http://en.wikipedia.org/wiki/MUSHRA) 
28 3 Steve Welburn
* "MUSHRAM"http://www.elec.qmul.ac.uk/digitalmusic/downloads/#mushram - Matlab interface for MUSHRA audio tests
29 1 Steve Welburn
30 1 Steve Welburn
Data involved includes:
31 1 Steve Welburn
* Software for the algorithm (which can be hosted on "Sound Software":http://www.soundsoftware.ac.uk)
32 1 Steve Welburn
* Original uncompressed audio
33 1 Steve Welburn
* Audio output of the new algorithm on the audio
34 1 Steve Welburn
* Audio output of existing algorithms on the same audio
35 1 Steve Welburn
* Documentation of the testing methodology
36 1 Steve Welburn
37 1 Steve Welburn
We note that for listening tests, the research may involve:
38 1 Steve Welburn
* Demographic details of participants
39 3 Steve Welburn
* Identifiable participants (Data Protection])
40 3 Steve Welburn
* Release forms by people taking part
41 1 Steve Welburn
42 1 Steve Welburn
and *will* involve:
43 1 Steve Welburn
* ethics-board approval
44 1 Steve Welburn
45 1 Steve Welburn
h3. Publication
46 1 Steve Welburn
47 1 Steve Welburn
Additionally, publication of results will require:
48 1 Steve Welburn
* Summarising the results
49 1 Steve Welburn
* Publishing the paper