Are there some open source alternatives to what Wit.ai does, i.e. mapping voice/text queries to actions by defining recognition patterns?
(Wit.ai has an open API, but source code is closed, making it inappropriate for education purposes.)
-
Is it English-only?– NemoOct 5, 2015 at 17:13
-
So far as I know, wit.ai is english-only, but i wouldn't mimd if thw alternative is multilingual. Or in some other language (preferably in latimic script for the sake of redability); if the approach to the prblem is probabilistic it isn't really cruical whether is english or not - it should be easy enough to modify some of the parameters, rarher than rewriting logics tiwd up to semantics of a specific language.m– smihaelOct 7, 2015 at 8:29
Add a comment
|
3 Answers
Baidu Research released yesterday the code for its speech recognition library using Connectionist Temporal Classification implemented with Torch.
Benchmarks from Gigaom:
answered Jan 18, 2016 at 17:43
-
Very cool, but CTC is just one part of a speech recognition system. You also need the rest of the network, and a huge amount of training data. Baidu used 10k hours of transcribed audio, expended to 100k hours with noise.– TimmmmFeb 12, 2016 at 13:47
-
1@Timmmm Good point. Some other pointers: Decent speech recognition software for Linux Feb 12, 2016 at 20:20
Snips.ai is an alternative now (2019). Open source and privacy respecting. Their code is on github
Edit:
their console is not open source anymore since it was aquired by sonos.
-
It served my purpose well, but they decided to close public access to the developer console forum.snips.ai/t/important-message-regarding-the-snips-console/…, which makes it less useful as some key parts of the code were closed source.– smihaelJan 21, 2020 at 7:55
There are no open source alternatives. Kaldi is probably the best open source speech recognition system, but it is extremely non-trivial to set up (I still haven't managed it actually). I've used Sphinx in the past but it is very low quality (think 2000's-era speech recognition - i.e. before it was any good).
As for the NLP side of things, what I call "intent recognition", I haven't looked into it that much, but Stanford's CoreNLP may be a good place to start. Again I expect it to be extremely non-trivial.
There are two similar closed-source alternatives - Nuance Mix, and Houndify, but as of now (Feb, 2016), they are both effectively closed betas (Houndify doesn't let you create custom commands yet, and Nuance hasn't opened their HTTP API, though you can access it via Android or iOS).
