We developed a solution for deaf and hard of hearing in
the form of an application, called UHO. The solution uses
a local speech recognition model to display real-time captions
of Slovenian speech that is being played on the device. For
this purpose, we used the speech corpus Artur 1.0 to train the
speech recognition models. We trained the models on the VEGA
supercomputer, where we achieved word error rates of 11.38%
for the larger base model and 15.19% for the smaller tiny model.
We ensured real-time execution by using appropriate model
inference backends and by using a token decoding optimization
approach that we call additive decoding. Results are promising,
indicating that real-time speech recognition can give comparable
results to recognition of full-length audio samples.
|
