OOV Detection Visualization Tool

>>win package<<         >>linux code<<

  • The NN-based OOV Detection visualized on a set of sample recordings is available
    as a package for Windows or alternatively in source code (requires Qt 4.5 or newer).
  • Shown are phone posterior probabilities of a specific recognizer (constrained by English language model) and phone posterior probabilities of a general phone recognizer using long temporal context (300 ms).
  • One-best phone hypotheses of both recognizers are highlighted in blue.
    At the bottom the recognition output is shown (word labels) and the corresponding word confidences.
  • The curve is a frame-level (10 ms) OOV probability estimated by the neural net. Several sample data sets are bundled with the package sporting different speakers with different origin, gender and age. Some examples of non-speech sounds are added, too.
    The windows package is self-contained (binary, no dependencies), the source code requires Nokia Qt 4.5 or newer to compile (MacOSx/Linux/Windows). 

Bild

>>win package<<         >>linux code<<