Currently sucessfully using pocketsphinx_continuous for word spotting. The use case is to spot company names in 1-2 hour podcasts. So far I've only been using -dict option to augment the dictionary with different pronunciations of company names.
I'd like to start using -lm option however when I generate lm file with my limited set of company names I get a lot of false positive matches, I used this tool. http://www.speech.cs.cmu.edu/tools. Also I get a lot of <sil> [NOISE][UM] etc in the output which I do not get when I omit the -lm option.
Does anyone have recommendations on how to use pocketshinx to optimally detect company name mentions in podcats?
Last edit: a k 2018-06-26
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thanks, I've looked into Kaldi before, but it's too complicated to implement as far as I'm concerened. I already have an aws server setup to process podcasts at scale with pocketsphinx, results have been generally good and I am able to update the dictioary with oov words which seems like would be difficult with Kaldi as well.
In your experience will kws be drastically improved with Kaldi? Do you have any suggestions on how to improve the process with pocketsphinx?
Setting up Kaldi to batch process 1000s of wav files like I have with pocketsphinx seems to be very too difficult.
Thank you.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Currently sucessfully using pocketsphinx_continuous for word spotting. The use case is to spot company names in 1-2 hour podcasts. So far I've only been using -dict option to augment the dictionary with different pronunciations of company names.
I'd like to start using -lm option however when I generate lm file with my limited set of company names I get a lot of false positive matches, I used this tool. http://www.speech.cs.cmu.edu/tools. Also I get a lot of <sil> [NOISE] [UM] etc in the output which I do not get when I omit the -lm option.
Does anyone have recommendations on how to use pocketshinx to optimally detect company name mentions in podcats?
Last edit: a k 2018-06-26
Modern keyword spotting is implemented in Kaldi:
https://github.com/kaldi-asr/kaldi/blob/master/egs/babel/s5d/local/kws_search.sh
Thanks, I've looked into Kaldi before, but it's too complicated to implement as far as I'm concerened. I already have an aws server setup to process podcasts at scale with pocketsphinx, results have been generally good and I am able to update the dictioary with oov words which seems like would be difficult with Kaldi as well.
In your experience will kws be drastically improved with Kaldi? Do you have any suggestions on how to improve the process with pocketsphinx?
Setting up Kaldi to batch process 1000s of wav files like I have with pocketsphinx seems to be very too difficult.
Thank you.