8 ms·

Actually I think most of the work has been done in academia --- certainly that's where the recent deep learning stuff has come from. So, I don't think the impor

Actually I think most of the work has been done in academia --- certainly that's where the recent deep learning stuff has come from. So, I don't think the important stuff is patented. (In general it's very hard to lock down ML improvements under patent. Once we can do it one way, and understand a little bit about what's working, we can usually replicate that performance with another technique.) The big problem for open source speech recognition is training data.

melling 12y ago
First, these are two different problems to solve. Voice recognition and deep learning are different fields. Is training really the issue for voice recognition? It has been a problem that has almost been solved for over a decade. Last year I saw this impressive use of Dragon Naturally Speaking for the PC, running in a VM on a Mac, that pretty much worked to code by voice. https://www.youtube.com/watch?v=8SkdfdXWYaI https://www.youtube.com/watch?v=8SkdfdXWYaI The developer mentioned that he didn't have any luck with Sphinx. Xah Lee summarized the talk here: http://ergoemacs.org/emacs/using_voice_to_code.html http://ergoemacs.org/emacs/using_voice_to_code.html
- walterbell 12y ago
  Thanks for that pointer to the Python library which integrates with Dragon/Nuance to enable arbitrary commands, https://pypi.python.org/pypi/dragonfly/ https://pypi.python.org/pypi/dragonfly/
- syllogism 12y ago
  ...Of course deep learning is not voice recognition. But the most recent advances in speech recognition have been from deep learning models, which have come from academia. The system in the video you link is single speaker, closed vocabulary. You need massive training data for multi-speaker, open vocabulary.
- bane 12y ago
  https://www.youtube.com/watch?v=KyLqUf4cdwc https://www.youtube.com/watch?v=KyLqUf4cdwc
  - melling 12y ago
    This video is not going to convince anyone to use voice recognition.
- IshKebab 12y ago
  I've tried to use sphinx, but the problem was lack of training data (you have to supply it yourself pretty much!). It did have some data that was supposed to recognise numbers, but it didn't work (I mean, it ran, but the recognition was awful even when it only had to pick between 10 options). Training is a huge issue for voice recognition. It's the only way Google and Apple have managed to take voice recognition from "works 80% of the time, but that is still bad enough to be totally usable" to "this actually works!". Maybe you don't remember how bad voice recognition was 10 years ago. To give you an idea how important it is, on OSX you have the option to download data to improve offline voice recognition. It's something like 500 MB. And that's the result of the training.
  - mikeash 12y ago
    I think there may be some confusion as to what "training" means. When it comes to voice recognition, it makes me (and I suspect others) think of the older software which required a user to read a bunch of text to it to train the software to your specific voice before it could do any kind of decent job understanding you. Now, everything is speaker-agnostic and works out of the box for anybody. Different kind of training.
walterbell 12y ago
> The big problem for open source speech recognition is training data. What exactly is needed for training - audio recordings with transcripts, human validation of recognized text? There are successful crowdsourced efforts for proofreading of OCR'ed text. Archive.org could host a CC-licensed archive of sound & transcripts. Recognition of the human voice is almost like writing, hopefully everyone could have access. Edit: how much disk space would be needed - TB or PB?
- kansface 12y ago
  The more data the better although the relationship isn't linear.
- albertzeyer 12y ago
  For example, the Switchboard corpus (300h, 8khz, transcribed audio) is about 16GB. That is a common size for LVCSR, and you need something around that area to get good performance (maybe minimum 100h). In academic papers by Google, they usually use their own private training data set, with e.g. 1900h. (E.g.: http://arxiv.org/pdf/1402.1128.pdf http://arxiv.org/pdf/1402.1128.pdf) Some crowdsourced effort to collect transcribed audio under a CC-licence would be great!
  - bainsfather 12y ago
    Maybe this? http://www.voxforge.org/home http://www.voxforge.org/home - "VoxForge was set up to collect transcribed speech for use with Free and Open Source Speech Recognition Engines (on Linux, Windows and Mac)." (caveat: I have not recorded on this from (any) of my machines - I don't have the right plugin apparently) Maybe also: https://librivox.org https://librivox.org - has audiobooks read by volunteers, plus the book text.

melling 12y ago

First, these are two different problems to solve. Voice recognition and deep learning are different fields. Is training really the issue for voice recognition? It has been a problem that has almost been solved for over a decade. Last year I saw this impressive use of Dragon Naturally Speaking for the PC, running in a VM on a Mac, that pretty much worked to code by voice. https://www.youtube.com/watch?v=8SkdfdXWYaI https://www.youtube.com/watch?v=8SkdfdXWYaI The developer mentioned that he didn't have any luck with Sphinx. Xah Lee summarized the talk here: http://ergoemacs.org/emacs/using_voice_to_code.html http://ergoemacs.org/emacs/using_voice_to_code.html

walterbell 12y ago
Thanks for that pointer to the Python library which integrates with Dragon/Nuance to enable arbitrary commands, https://pypi.python.org/pypi/dragonfly/ https://pypi.python.org/pypi/dragonfly/
syllogism 12y ago
...Of course deep learning is not voice recognition. But the most recent advances in speech recognition have been from deep learning models, which have come from academia. The system in the video you link is single speaker, closed vocabulary. You need massive training data for multi-speaker, open vocabulary.
bane 12y ago
https://www.youtube.com/watch?v=KyLqUf4cdwc https://www.youtube.com/watch?v=KyLqUf4cdwc
- melling 12y ago
  This video is not going to convince anyone to use voice recognition.
IshKebab 12y ago
I've tried to use sphinx, but the problem was lack of training data (you have to supply it yourself pretty much!). It did have some data that was supposed to recognise numbers, but it didn't work (I mean, it ran, but the recognition was awful even when it only had to pick between 10 options). Training is a huge issue for voice recognition. It's the only way Google and Apple have managed to take voice recognition from "works 80% of the time, but that is still bad enough to be totally usable" to "this actually works!". Maybe you don't remember how bad voice recognition was 10 years ago. To give you an idea how important it is, on OSX you have the option to download data to improve offline voice recognition. It's something like 500 MB. And that's the result of the training.
- mikeash 12y ago
  I think there may be some confusion as to what "training" means. When it comes to voice recognition, it makes me (and I suspect others) think of the older software which required a user to read a bunch of text to it to train the software to your specific voice before it could do any kind of decent job understanding you. Now, everything is speaker-agnostic and works out of the box for anybody. Different kind of training.

walterbell 12y ago

> The big problem for open source speech recognition is training data. What exactly is needed for training - audio recordings with transcripts, human validation of recognized text? There are successful crowdsourced efforts for proofreading of OCR'ed text. Archive.org could host a CC-licensed archive of sound & transcripts. Recognition of the human voice is almost like writing, hopefully everyone could have access. Edit: how much disk space would be needed - TB or PB?

kansface 12y ago
The more data the better although the relationship isn't linear.
albertzeyer 12y ago
For example, the Switchboard corpus (300h, 8khz, transcribed audio) is about 16GB. That is a common size for LVCSR, and you need something around that area to get good performance (maybe minimum 100h). In academic papers by Google, they usually use their own private training data set, with e.g. 1900h. (E.g.: http://arxiv.org/pdf/1402.1128.pdf http://arxiv.org/pdf/1402.1128.pdf) Some crowdsourced effort to collect transcribed audio under a CC-licence would be great!
- bainsfather 12y ago
  Maybe this? http://www.voxforge.org/home http://www.voxforge.org/home - "VoxForge was set up to collect transcribed speech for use with Free and Open Source Speech Recognition Engines (on Linux, Windows and Mac)." (caveat: I have not recorded on this from (any) of my machines - I don't have the right plugin apparently) Maybe also: https://librivox.org https://librivox.org - has audiobooks read by volunteers, plus the book text.