This month was fairly busy in terms of FreeDV. Besides going to Pacificon and Zero Retries Digital Conference, a lot of work was done in terms of bugfixes and testing which ultimately resulted in a 2.4.1 bugfix release going out at the end of the month. Many of these bugfixes were part of an effort to improve automated testing in the upcoming FreeDV 3.0 release, much of that directed by Claude. In the process of writing tests, it discovered many bugs that would likely have taken quite a while to discover manually if at all.
In addition, because of this automated test effort, we now have automated unit and GUI tests. The latter in particular exercise as much of the user interface as possible thanks to a Python library and a TCP/IP link to the FreeDV application that it controls. This lets us have code coverage exceeding 90%, along with a way to make sure future changes do not break anything existing.
For RADE V2, some improvements with callsign decode were made that take advantage of the fact that the same callsign is repeatedly transmitted. This allows the callsign to be decoded far more often and at lower SNRs, meaning that your transmission is far more likely to be reported by others to FreeDV Reporter and elsewhere.
Finally, after 2.4.1 released, I began (with Claude’s help) to implement a small machine learning model that takes 8 kHz audio (most commonly with Bluetooth headsets that don’t fully support the mSBC codec) and effectively “fills in” the missing features for the RADE encoder. This allows the receiver to hear a transmission that has some of the higher audio frequencies added back in, improving clarity in this (admittedly limited) scenario. Some audio samples below:
Original audio:
Audio after RADE encode/decode:
Audio after downsampling to 8 kHz, going through new model and passing through RADE encode/decode:
Some adversarial training was also suggested and performed by Claude, which improved the results further. Examples from before and after such training:
This is currently split into two pull requests: freedv-backend for the actual ML model and freedv-gui to enable it when the mic device’s sample rate is low enough.
More information can be found in the commit history below:
(Note that all commit logs above were generated with the following command line:)
git log --author="member@email" --after "Month 1, 2026" --before "Month 31, 2026 23:59:59" --all
