How to build really robust speech to text technology

Posted on December 19, 2007
Filed Under Business, Google |

01.pngBoy, this is interesting:

The reason we really did it is because we need to build a great speech-to-text model … that we can use for all kinds of different things, including video search.

Google developed Goog 411 to train their speech to text technology. Via Waxy who remarked it is a neat idea but seems kinda inefficient…I dunno, seems like a super smart way to get access to a huge set of data to work with, and if it results in a way superior technology then the ramifications would be huge.

Comments

3 Responses to “How to build really robust speech to text technology”

  1. Spudooli on December 19th, 2007 7:33 pm

    How evil do you think they can become if they can convert speech to text? Just about everything you can imagine will be theirs to own.
    Relevant and contextual ads interjected into your conversations while you are using their new 700MHz wireless on your Google supplied free VOIP?

  2. Andy Baio on December 19th, 2007 9:24 pm

    I knew I should’ve tried to clarify my meaning there… I was trying to say that there must be better existing sources for parallel text/speech out there already. For example, any closed captioned dialogue with spoken audio.

    I’m suspicious about Marissa Mayer’s claim that it was a primary reason why GOOG-411 was built. It seems like an awful lot of work just to build up a corpus, but maybe I’m totally wrong. Regardless, it’s very interesting.

  3. Charles on December 20th, 2007 8:07 am

    Dave, exactly. I’m thinking online video as the low hanging fruit, and then radio and TV(?).

    Hey Andy, good point. It does seem like an awful lot of effort for “free” data. Even if the service pays for itself I can’t imagine the money would be an barrier if the data was available elsewhere.

    So maybe that is the issue…I can’t myself think of another place to find that many samples (millions), or atleast somewhere that would sell them to Google.

    Add the benefit of having a purpose built data set rather than re-purposing another and maybe it makes sense. The implications of getting this stuff really right are definitely huge.

Leave a Reply