[asterisk-biz] Asterisk and Yext.com's (on the fly) call transcription..

[Ritesh A]

Peter,

Thanks for the insights. I just tried out Twilio's Speech to Text API and
played the same message as Yext Demo. It worked out pretty well so I am very
impressed. It's likely that Twilio uses SpinVox in the back-end anyway.. By
the way, I think Yext records the call and sends the recorded file for
transcription. This saves the hassle of mixing and transcribing two
different audio files.

Yes, Yext has a fairly impressive solution indeed and I could think of so
many applications that are vertical specific (e.g. Automotive). I could
imagine Yext building a lot of library relevant to each category (e.g.
Doctor, Mechanic, Finance etc) and then doing a close/guess association of
the returned text to white/blacklist certain terms and then splitting them
into different sub-categories (e.g. Year built, Make/Model etc). This is
hard work (hence barrier to entry) and they seem to have raised $17M in
recent round in addition to running a (self asserted) $20M and growing
business. Hard to play the catch up game...

Another interesting fact is that Yext solution would improve with time much
like wine. Given the call/customer flow and lack of competition, Yext could
continue to improve the keyword filtering and categorizing over time
creating a fairly accurate solution.

Ritesh



On Wed, Sep 30, 2009 at 9:31 AM, Peter Beckman <beckman at angryox.com> wrote:

> On Wed, 30 Sep 2009, Ritesh A wrote:
>
>  Folks,
>>
>> Does anyone have a clue on how Yext transcribes the calls (almost) on the
>> fly?
>> I am pretty sure they are using asterisk at the heart and perhaps using
>> some
>> neat APIs for call transcription.
>> If you are know of any outfit there that provides similar APIs, please
>> ping
>> me...
>>
>> If anyone is interested in learning more about Yext, here is the video
>> demo
>> of their interesting demo at TC50
>>
>> http://www.ustream.tv/recorded/2163590
>>
>
>  I am extremely impressed with Yext.  Thanks for the link and your post!
>
>  Here's my take on how they do it.
>
>  * recording both ends of the call separately (Asterisk can do this)
>  * pass both audio files to a voice processing API (SpinVox)
>  * integrate the relating text into a conversation flow (very cool, not
>   sure how, depends on the API)
>  * parse resulting text into keywords matching categories (maybe scoring
>   them somehow), then if the score is high enough, using a higher level
>   matching (if a year (2007), make (Porche) and model (911) are mentioned,
>   put them together in the right order and put it in the inbox,
>   searchable).
>
>  Both the merging of the two sides of conversation and The keyword matching
>  is where they really did some cool stuff.  While I can think of how they
>  did it in theory, it probably took a lot of work to pull that off.
>  "Automotive" is likely a set of related keywords, and then another set of
>  keyword combinations.  One of the combos combine the matches on keywords
>  found, then matches those to "[year] [make] [model]" as a relevant
>  keyphrase if Automotive is selected.
>
> Beckman
> ---------------------------------------------------------------------------
> Peter Beckman                                                  Internet Guy
> beckman at angryox.com
> http://www.angryox.com/
> ---------------------------------------------------------------------------
> _______________________________________________
> --Bandwidth and Colocation Provided by http://www.api-digital.com--
>
> AstriCon 2009 - October 13 - 15 Phoenix, Arizona
> Register Now: http://www.astricon.net
>
> asterisk-biz mailing list
> To UNSUBSCRIBE or update options visit:
>   http://lists.digium.com/mailman/listinfo/asterisk-biz
> _______________________________________________
> --Bandwidth and Colocation Provided by http://www.api-digital.com--
>
> AstriCon 2009 - October 13 - 15 Phoenix, Arizona
> Register Now: http://www.astricon.net
>
> asterisk-biz mailing list
> To UNSUBSCRIBE or update options visit:
>   http://lists.digium.com/mailman/listinfo/asterisk-biz
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.digium.com/pipermail/asterisk-biz/attachments/20090930/2257add1/attachment.htm