This version is a major update from v3.1.0.
- Introduces a deep lattice network model for speech mode, which improves predictions over the exponential model by using multidimensional partially monotonic mapping of similarity to quality. It is enabled by default (and configurable with
–use_lattice_model). There is no lattice model support for audio mode yet, but we will be looking at that for the version. - Adds support for python via pybind11 (see the test in the python subdirectory for an example). This is supported on Linux and Mac. (Windows only supports C++, Windows developers are welcome to help).
- Various bug fixes and code improvements, including DSP fixes, which will slightly change the predictions for all models compared to previous audio and exponential speech models.