Just letting you guys know, Im actually really happy at the moment :) Im always posting when Im sad, so I might as well post when Im happy too! ^_^ Anyway, Ive come up with a great new idea. You know how they have speech to text and image to text and all other sorts of algorithms that convert analog data from the environment (e.g. sounds, pictures, etc) which contains natural language and decode it into digital text? Well I want to make such an algorithm... FOR SIGN LANGUAGE!!! A program where you sign in front of a camera and it watches you sign and as you sign it works out what signs you are doing. (And then as an addition step it can run it through a translation algorithm into a spoken language like English, but this is just a secondary step, and is no different to machine translation as it already exists). I dont really know much about computer science or actual proper programming languages and I have no skills or experience with machine translation or these sorts of algorithms (I dont even know what theyre called... analog recognition algorithms?), but I *do* have experience/interest/skills with 1) Linguistics 2) Mathematics (and also some basic programming experience, but not with any actual like standard languages, not yet anyways, but I am so willingly to learn programming and computer science) 3) Auslan 4) Im interesting in SignWriting, a method for writing and recording signs with all their non-linearity and intricasy. I mean, how many people out there are interested / have knowledge about in Linguistics AND a sign language AND are mathematically/programmingly minded? I just need to find some people who know more about the algorithmic side of things than I do, and yeah.......... this is something Id really like to do. This would be a great cause to dedicate my life to... I would love to do field linguistics with Auslan and work out the linguistics of the language... do machine translation and algorithms... work on the programming and mathematics and linguistics of it all... this would be such a great way to combine all my different interests and skills and create a product that visibly tangibly helps people and makes their quality of life better. A whole part of these Auslan courses is learning about the Deaf culture and community, and it would be so awesome if I could make something that would bridge the communication gap between them and hearing people. Also, on the topic of speech/image to text converters, I really wish that they used the IPA as an intermediate step. I think it would be really good if there were each of the following algorithms that existed independently: - text to IPA (phonemic) - IPA (phonemic) to IPA (phonetic) - IPA (phonetic) to speech - speech to IPA (phonetic) - IPA (phonetic) to IPA (phonemic) - IPA (phonemic) to text - image to text So like, youd say hello into the microphone, and the speech to IPA (phonetic) algorithm would get the sound waves and interpret them and abstract out the sequence of IPA symbols which best represents what was spoken. Then the IPA (phonetic) to IPA (phonemic) algorithm would get the sequence of phonetic IPA symbols and, with knowledge of English phonology (and possibly having been trained to your particular accent), will try to match the sequence of IPA phones to a sequence of English *phonemes*. Then, the IPA (phonemic) to text algorithm would get this sequence of English phonemes and, using its knowledge of English spelling and a list of English words, match the phoneme sequence to its best guess of the intended word. The only problem with this is that while the match between phone and phoneme can be trained and adjusted using learning algorithms, each word will have to have a corresponding *phonemic transcription* (or at least a range of them) preprogrammed into it. And these phonemes are discrete. And in fact, the whole set of phonemes has to be preprogrammed. While this makes the system very elegant, it does make it awfully..... predefined. Its sort of like the question of machine translation... do we program it with rules or do we just give it a corpus and tell it to work them out itself? Modern machine translation like Google Translate does the latter, and it works much better than the former, although it is more unpredictable when the corpus is not big enough. Maybe we should skip the whole IPA (phonemic) stage. Maybe we should just have a fixed algorithm to analyse sound waves and abstract the exact narrow phonetic IPA transcription, but the association between the exact narrow phonetic IPA transcription and each word is done via machine learning. But yeah. There are so many text to speech programs you can get, I just wish there was an IPA to speech program...
Posted on: Tue, 11 Nov 2014 20:40:32 +0000
Trending Topics
Recently Viewed Topics
© 2015