KeenASR

Hi @sagarmodi We currently support English. With our SDK you as a developer have to specify what the engine is listening to; for small to medium size vocabulary tasks (up to a few thousand words) this can be done via API, by providing a list of phrases. For a given list of phrases we will build a small language model and decoding graph, which can then used for recognition. For larger vocabulary, language model (and the corresponding decoding graph) need to be built ahead of time using large corpus of text data. At this time we would need to help a bit with decoding graph creation (you could create ARPA language models), but pretty soon we'll expose some tools so developers can do this without our involvement. When comparing to pocektsphinx or any Gaussian Mixture Model solution, you can expect to see at least 2x improvement in terms of Word Error Rates (e..g if you are getting 10% WER with sphinx, you'll like see 5% or less with our SDK). One of our customers has done comparison of "largish" dictation (about 40,000 words, so not very large), in a niche domain. They happen to have text data to train the language model and they used this language model to test our SDK, as well as IBM and Microsoft cloud solutions (they both provide ways to use custom language models). All three solutions were performing at about 95% accuracy (5% error rate), with our solution the only one that was not cloud based. Depending on your use case, data availability, etc. the best approach is to create a small test set that's relevant for your use case (so, it's as authentic as possible), then run different SDKs against the test set in a batch mode. (this can also be used to fine tune the configuration). Feel free to email me (ogi at keenresearch dot com) if you have any questions.