4
$\begingroup$

One could train a speech recognizer with words or phonemes. I guess that they both have their pros and cons:

  • with words there would be a lower error rate as phonemes can be put together building a non-existing word (this case would be impossible with word training)
  • using words would require a larger database and more space and processing time
  • phonemes are way more general: with all phonemes you can build all existing words, but with all the words in the database you can't recognize all the words in the language

I'm wondering if there is a reason to prefer one over another, or if in some cases one is better than the other. Do both methods have practical applications? Which ones? What are their pros and cons (apart from the ones I named, which may be wrong so feel free to correct me)?

| improve this question | |
$\endgroup$
  • 1
    $\begingroup$ You got it all correctly, not sure what your question is about. $\endgroup$ – Nikolay Shmyrev Jul 24 '17 at 10:39
  • $\begingroup$ @NikolayShmyrev I wanted to know if there were any other differences, and in which practical applications one would be better than than the other. $\endgroup$ – Tendero Jul 24 '17 at 14:40
  • $\begingroup$ As I wrote you've got main differences correctly. Phonetic recognizers are used in callcenter products where you have many unusual names and need to search for them, Nexidia is using them actively. If you know the vocabulary beforehand you can use word recognition system, practically every other serious system is based on words. There are also end-to-end recognizers by Baidu which recognize letters instead of words or phonemes, but they are not yet practical. $\endgroup$ – Nikolay Shmyrev Jul 24 '17 at 14:53
  • $\begingroup$ @NikolayShmyrev If you write an answer containing this information, I can accept it to close this post. $\endgroup$ – Tendero Mar 14 '18 at 19:51
1
$\begingroup$

Most modern speech recognition systems are based on Hidden Markov Models. These are probabilistic models that require a previous training. So here is simple comparison:

  • By using phonemes you have a reduced number of possibilities (around 50 depending on the language), you get a lot of data to train each phoneme model, by combining several models you can have word models.

  • By using word models since the beginning, you would need to have thousands of models to train and probably not enough data.

If you are working on a very simple recognition system, with a very limited set of words, for example, recognition of numbers, then maybe you can use words models that don't require HMMs.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.