overlapFactor
The proportion of overlap that the feature-extraction session uses to analyze two consecutive windows in the audio data.
Declaration
var overlapFactor: DoubleDiscussion
The overlap factor — which must be in the range [0.0, 1.0) — affects how much audio data the feature- extraction session analyzes in each file. Sessions with smaller overlap factors read fewer audio data samples and finish in less time but may compromise the model’s prediction accuracy. Sessions with larger overlap factors read more audio samples for each file, which increases the session’s training data and processing time. The additional training data can improve a sound classifier’s accuracy; however, it may only be a modest improvement that isn’t worth the extra processing time.
The feature-extraction session uses the expression (featureExtractionTimeWindowSize * (1.0 - overlapFactor)) to determine the how much to step (advance) in time between samples.
Window size | Overlap factor | Step time |
|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
For example, a session that’s analyzing a 5-second audio file with a window size of 1.0 and an overlap factor of 0.0 samples the audio five times. The time offsets for those samples are: 0.0, 1.0, 2.0, 3.0, and 4.0.
Another session with a window size of 1.0 and an overlap factor of 0.5 samples the same audio file 10 times at half-second intervals. Unlike the first session, this session samples each portion of audio data twice, except for the first and final half-second.
A third session with a window size of 1.0 and an overlap factor of 0.75 samples the same 5-second audio file 20 times at quarter-second intervals. This third session samples most of the audio portions four times; the session samples the intervals near the ends one, two, or three times.