For example, knowing that “I'd like a nice cup of ” preceded the word I'm trying to process now, helps my algorithm determine that there's a higher probability of the word being “tea” (if it's a short, one-syllable word with certain phonemes in it), or “coffee” (if it's a disyllabic word….) – anyway – you get the idea.

We humans do the same thing. If you fire a word at me, chopped out of an audio stream, it's often very difficult to determine what that word is. “Tea” is a good example: depending on the audio environment, quality of the line etc., “tea” can sound like “tea” (correct), “tee” (different spelling – context – meet you at the first tee), “me”, “be” and so on.

So,

- if I were running a service that offered voice-to-text as part of the solution (and I do – I work for Dial2Do)
- and I wanted to ensure good overall transcription quality (and I do)
- then – when I offer audio to humans to be corrected, it behoves me to offer the whole stream, because that way, they have the whole context, and therefore the best chance of transcribing correctly

Even if you aren't a speech recognition nerd, common sense tells us this is so.

As a result, I don't believe SpinVox do it this way. I think they offer the wider audio message to help the tranbscribers do their job. It. Just. Makes. Sense.

I also don't think there's anything wrong with that (Nuance, Dial2Do, PhoneTag and others all do likewise).

And lastly – I still have severe envy for the SpinVox Stand at CTIA and MWC. Awesome!

Cheers, Sean

That's simply not true and a complete and utter fabrication.

SpinVox does employ human agents to assist the main system when it *doesn't* know a word or sound. These are illustrated in your messages by underscores and/or questions marks like so:

___ or (?)

When this happens, the sound itself is cut *out* of the message, detached from any numbers and/or recognisable features and then put into a queue.
Further down the line (normally within a couple of weeks), these words/sounds are then listened to and converted manually into the system so as to not get it wrong in the future.

When required (then and only then), this is done post-event and never, ever 'in-situ'

I hope this clarifies things for you,

Cheers,

James

PS We get these accusations from either disgruntled ex-employees and/or bitter rivals from time to time. Take it as a compliment as they only ever seem to appear on the more popular blog posts about the SpinVox service.

http://twitpic.com/2am96

http://voicemailtranscription.blogspot.com/2008...

SpinVox is currently rolling out on the Telstra network in Australia.
Hope that is of use to you!

Cheers,

James