Questions about keyphrase spotting

[Otherend]

Hello everyone,

I have several questions about spotting keyphrases with pocketsphinx:

1) What is the best way to chose the threshold value for a keyphrase? (there will be many keyphrases)

  a: recording several persons saying the keyphrase
  b: recording several persons saying the keyphrase in a sentence
  c: inserting several records of the keyphrase in a long audio speech at specific times (example every minutes in a 2-hour speech)
  d: if it is a short word chose a higher threshold, if not a lower threshold
  e: other ways (which ones?)

2) How many different speakers should record the keyphrase?

3) Is keyphrase spotting becoming less accurate when the number of keyphrases to be analysed increase?

4) And is it becoming slower? With 100 keyphrases? 1000?

5) I am thinking about automating the choice of the threshold. For example, it will try different threshold values for a given keyphrase, and I will chose the level of acceptance I want. Is it a good idea? (I just don't want to lose time by having to do it manually)

6) Is the idea in 5) already coded by someone?

Thank you for your consideration,
Otherend

[Nickolay V. Shmyrev]

1) What is the best way to chose the threshold value for a keyphrase? (there will be many keyphrases)

This issue is covered in http://cmusphinx.sourceforge.net/wiki/tutoriallm#keyword_lists

2) How many different speakers should record the keyphrase?

Testing allows you to figure out how system will behave in a production. So you need to try to reproduce the production environment as close as possible. If many speakers will use the system, you need many test speakers as well. However, single speaker also gives a reasonable approximation for threshold.

3) Is keyphrase spotting becoming less accurate when the number of keyphrases to be analysed increase?

Yes

4) And is it becoming slower? With 100 keyphrases? 1000?

Yes

6) Is the idea in 5) already coded by someone?

No

[Otherend]

Thank you !