Supposing to use Festival to build a speech synthesis system and HTK for a speech recognition system. The language is a Native American language.
Which system would be harder to accomplish?
-
3Recognition is harder. Synthesis flows along fairly predictable set of tasks. Even synthesis techniques that are 30 years old produce understandable speech. New research is about making synthesis sound more natural. For recognition, you need a lot of training data, you might need to customize it for specific domains, accents, etc.– prash ♦Mar 28, 2014 at 16:08
-
1Recognition is much harder. At least if you want it to work with more than one speaker; especially if you can't constrain the context. And with a Native American language you're going to have some interesting challenges even for synthesis.– jlawlerMar 28, 2014 at 19:02
-
1@prash domains & accents in recognition are exactly problems that are very amenable to just throwing more data at them. However, “accomplish" is quite ill-defined here. For recognition, you would have to specify what kind of error rate would be acceptable, and for synthesis, it might be quite hard to even agree on the applicable quality metrics. How natural does the intonation have to be? How many phonetic errors are acceptable? What text normalization domains need to be handled ? (Does the system need to handle reading a particle physics paper written in Cherokee?)– microtherionMar 28, 2014 at 19:27
-
1Recognition seems clearly harder, but maybe your choice should not be dictated by difficulty of the task, but by the purpose of doing it, since you have a choice. It is like asking for a choice between swiming and tennis ... it may depend on your purpose more than on the difficulty of either sport.– babouMar 29, 2014 at 5:42
Add a comment
|
1 Answer
Without question, it is an order of magnitude more difficult to implement speech recognition.
You can do passable speech synthesis knowing nothing about context, grammar, etc. You will get some issues with homographs but other than that you will get recognizable speech. Now, it will be greatly lacking in prosody and stress, so you need to do some language analysis to be successful but it's all doable and been done. There are toolkits out there that will help you build speech synthesis for any language even without sampling.
But the speech signal is just too noise for speech recognition to be done with any level of success without some understanding of language. This is mostly done stochastically these days, so the recognizer will guess what is likely to come next. So unlike with synthesis you need huge amounts of speech data to process. Again there are open source toolkits that will help you get started, but the amount of work you have to do for a new language is huge and the results will never be as impressive as with speech synthesis given the same amount of work.
-
Generation is probably harder for character-based languages with many homographs. Mar 30, 2014 at 8:38