zplane -- music processing and analysis technology

Sub Navigation

• 3.1: music creation
• 3.2: DJ
• 3.3: music analysis
  • 3.3.1: [aufTAKT]
  • 3.3.2: metering
  • 3.3.3: [museg]

    • Current page is 3.3.3.1: description

    • 3.3.3.2: specs
  • 3.3.5: [tONaRT]
• 3.4: broadcast
• 3.5: speech

The [museg] SDK is able to segment audio input data into segments containing music and segments containing speech and can be used e.g. for broadcast monitoring etc.

The SDK consists of two parts: the training part and the classification/segmentation part. With an individual selection of training files, it is easily possible to adapt the algorithm to special use cases and thus increase the classification performance significantly.

• Feature Extraction: a set of meaningful and characteristic parameters is extracted from the audio samples in a sliding time window

• Classification: for each time window, the extracted feature vector is reduced/transformed to a lower dimensional space and a decision is met to what class this specific observation belongs to

• Segmentation: the classification result is post-processed with a path finding algorithm that tries to determine the path with the highest overall probability for the whole audio file resp. all time windows. This stage allows especially to avoid short segments with low probability and thus smoothes out the overall result

The training system contains the same feature extraction stage as the classification, and afterwards iteratively optimizes the dimensionality reduction transformation matrix for these feature to optimize the classification results.

references

[museg] segmentation engine is successfully being used by:

• SoundAware MusicTracker

pricing & licensing