Demo doesn't work

[codersquare-com]

Hi, Great stuff. But I tried the digits and OSes on the demo page but it doesn't work. It often recognizes one as four , and many redundant digits I thought I'd just let you know. This kinda prevents me from using the library when the demo doesn't work well

[syl22-00]

Thanks for giving it a try.

Speech recognition accuracy and ways to improve it is a vast topic. To start with, Pocketsphinx.js should give you the same accuracy as PocketSphinx, and strategies to obtain good performance are the same. For instance, the library ships with a very small and rough acoustic model from which you should not expect too much. To build a serious speech app from open source software you should consider training your own acoustic model and carefully build language models.

A few pointers:

• An entry in pocketsphinx.js wiki about recognition accuracy.
• An entry in pocketsphinx.js README about recognition accuracy and performance.
• CMU Sphinx documentation page which contains extensive docs on building language and acoustic models.

I hope that helps.

[codersquare-com]

Thanks for ur explanation, looks like I have a lot to learn first

[syl22-00]

Yes, building a decent speech app is tough, that's why there are not that many of them...

One tip: confusions are much less likely with longer words. That's why there are less errors in the Cities demo.

[jergason]

Is there a place to get standard language and acoustic models to avoid having to construct your own?

[syl22-00]

Here are a few pointers:

• In Pocketsphinx source code, there are a few acoustic models, in the model/hmm/ folder. However, they might not be suitable directly (hub4 is in 8kHz, tidigits uses its own phoneme set).
• Voxforge has a few pre-trained acoustic models.
• Slightly more involved, you can follow the model training tutorial on the RM1 set.

[robertjchristian]

Very cool. Curious as to whether this could be used practically, ie one hour of spoken English conversation.