My question regards performing keyword spotting for custom keywords and justifying the use of keyword spotting models instead of speech recognition.
I have been doing some searching around Keyword Spotting and I realized there is not so much work out there. Probably the most common dataset I have found people using is the Speech Commands Dataset. However, this dataset has only 30 keywords.
If I want keyword spotting for my own custom application, then to the best of my knowledge I need either a pre-trained model or my own data to train a model on. However, to the best of my knowledge, there is no model pre-trained on a dataset with a large enough set of keywords that is likely to cover a lot of applications. Correct me if I am wrong in this.
I have come to the conclusion that I need to train my own model and the only two ways I could train models on custom keywords is to get that data myself, either by crowdsourcing or by performing speech recognition on large datasets, picking up segments which include the words of interest and then doing some manual work to check if these segments truly include the keywords I want. Does someone think that this would be a good or bad idea and why?
Lastly, why would I even bother going the keyword detection route and not just use a speech recognition model that will recognize the words a human speaks and see if any of them match my keyword? Is the performance that much better with keyword detection?