3 ms·
Looks like this isn't based on OpenAI's whisper, but on a model trained with RNNT loss using K2/icefall (https://github.com/k2-fsa/icefall https://github.com/k2
by woodson 4y ago
Looks like this isn't based on OpenAI's whisper, but on a model trained with RNNT loss using K2/icefall (https://github.com/k2-fsa/icefall https://github.com/k2-fsa/icefall; K2 is next generation of Kaldi). As such, it should in principle support customized vocabulary and LM for beam search decoding.