Is online decoding part of the (future) objectives of SpeechBrain? I am thinking about replicating the classical hybrid setup (acoustic model, lexicon, language model) in SpeechBrain to facilitate my research, but at the end of the day I want to make the models accessible through a speech-to-text API. Will that require translating insights back to (for example) Kaldi?
Hi! It is not available yet, but it is on the short-term list. We have to see what direction we want to follow here as K2 is coming. If K2 is nicely done, commented, and packaged, we could benefit from their expertise and simply provide a nicely explained interface. In the meanwhile, we need to think about how to achieve online decoding with the current E2E models. HMM-DNN ASR is also on the to-do list, but we need more peoples on it I guess