Processing VHF/UHF data, commonly found in aviation, maritime communications,
and emergency services, presents significant challenges. In addition to the
acoustic noise in the original signal, the transmitted signal is impacted by
various factors associated with the VHF/UHF communication channel. Notable
issues include multipath fading, interference from adjacent signals, and
signal loss caused by adverse weather conditions.
To be effective, the speech recognition model must account for these sources of
additive noise and distortion. One way to develop a representative model for
this data is to train it solely on properly annotated VHF/UHF data specific to
the targeted tasks. However, the high cost of data collection and annotation
makes this approach not very practical. An alternative is to use a representative
model of the VHF/UHF channel and apply it to existing training corpora. In
practice, these two approaches can be combined to strike a balance between
cost and performance.
We have developed effective VHF/UHF models to process both ATC data and military voice report data.
In 2018, Vocapia already demonstrated the effectiveness of its solution
which was ranked first for both the speech recognition and call sign detection
tasks in the Airbus ATC challenge.