Version used in paper

This is the same as v0.3.2, but retagged for citation in the paper on the syllable detector technique and performance.

Add support for Arduino output

Note: This release fixes a few important bugs since the Swift 3 migration.

In addition to many behind the scenes changes (including improvements to code organization, better memory management and updating to Swift 3 syntax), this update adds user interface options to enable using an Arduino to generate TTL pulses following syllable detection. The Arduino interface has lower latency than the Audio interface.

Additional optimizations and further network support

This update includes support for new processing functions, such as ``mapstd` as well as the equivalent of L2 and zscore normalization implemented as input processing functions. In addition, multiple input and output processing functions can be chained together, increasing flexibility of the networks that can be carried over.

Additionally, the code no longer tries to tweak FFT parameters for different sampling rates. Instead, it implements a very crude resampling algorithm to do linear interpolation in order to resample the incoming audio.

Finally, there are small tweaks to the naming conventions in the text file.

Support for new network types and better error detection

This release improves error detection during network conversion, and adds support for autoencoder networks, via the logsig and satlin transfer functions.

Better device selection and input/output meters

This release will automatically refresh devices on the device selection menu, and includes a working level-in indicator (although scaling likely needs some tweaking) and improves the level-out meter (no longer under sampling the neural network outputs).

Initial release of the syllable detector

An initial version of a syllable detector application that uses Matlab trained neural networks to perform low-latency syllable detection for identifying syllables in bird songs. The application can run multiple processors and run multiple neural networks on different channels. There are still some known limitations (see the issues section).