Importing and Reviewing Training Data

To start a new recognizer, use the "File" menu and select "New Recognizer". Alternatively, you can also open a previously generated recognizer for editing by opening the Song Scope Recognizer file (.ssr) directly by selecting "Open..." from the "File" menu.

The recognizer window looks just like the windows used to view audio files, except that the waveform and spectrogram plots are pushed over to the right side of the window to make room for a recognizer control panel on the left side.

Use the "File" menu and select "Import Notations" to import files containing training data. These can be the .wav audio files that you have already annotated, or you can open the corresponding Song Scope Notation files (.ssn) in the SongScopeNotes subdirectory directly.

The recognizer control panel shows each imported vocalization sorted by class and subclass in a tree structure. By default, each new loaded vocalization is selected as indicated by a check box next to the vocalization line. Each vocalization indicates the source recording file from which it originated, the time index and duration of the vocalization, and the Id as specified when making the annotation, or automatically assigned uniquely to each recording by Song Scope.

You can click on each individual vocalization line to view the vocalization in the spectrogram and/or waveform plots. Note that by default, the recognizer displays the waveform plot using Logarithmic Scale with Signal Detection, and the spectrogram plot using Logarithmic Scale with Signal Normalization. These views are important because they reflect how Song Scope will "see" the visualizations for building models.

You can double click a vocalization line to toggle between selecting and unselecting it for inclusion in generating a recognizer. You can also double click a class or subclass (or "All Classes") to select or unselect all of the vocalizations contained within.

You can right click (or Command-C for Mac OS X) on the tree to copy the annotation list to the clipboard, and then paste it into a spreadsheet or text file. This feature is handy for building reports to list the training data used.

You should adjust the settings described in the sections on Feature Reduction and Signal Detection Controls for optimum results and review each of the included vocalizations to make sure they are representative of the vocalization you are interested in and are not corrupted by noise that could contaminate the recognizer.

Previous Page Next Page