Re: [HTML Speech] speech resource specification requirement

On Sep 9, 2010, at 1:20 PM, Satish Sampath wrote:

>> Consider the following use case, a company, let's call them ACME Corp.
>> wants to put out a speech enabled web page that allows users to search
>> for their various products and services using voice. As part of their development
>> effort, they build a language model that supports this task.  With HTML+Speech
>> allowing specification of a speech resource on the network, they can serve the
>> same speech enabled page to all desktop and mobile browsers
>> supporting the standard.
> 
> Wouldn't it be sufficient to build a grammar based on the ACME product
> list than a whole language model? After all ACME corp may not have the
> resources or time to train with all possible voice variants and may
> alienate users in the process. Whereas a UA which supports speech
> recognition has the incentive to do it well enough to work for all web
> pages and use cases.
> 

Not if they want performance. If they want it to work well they
(or their vendor) will build a statistical language model. 
As users make queries they'll collect and transcribe that data
that train better models. Without the ability to specify 
a speech resource on the network they have no way to do this. 

>> We now have a situation where users will have a different
>> experience using speech input depending the browser, differing accuracy,
>> possible differences in tokenization and normalization.
> 
> This would already be the case if the UA decides to select a local
> recognizer instead of remote, per Eric's earlier proposal (whether it
> is because the local recognizer is more tuned to his voice or for
> bandwidth/speed reasons). I think we should let the UA decide the best
> configuration for the user rather than the web developer, as other
> APIs have done.
> 

To support broad range of use cases, you really need to support
both local and remote recognition, and user specified and 
developer specified models.  The requirement is not that users or developers
must use a remote model or that they must use a custom model,
rather that they have the option. If no remote resource is specified then
the default can be to be use the default browser associated model. 

best
Michael



> --
> Cheers
> Satish

Received on Thursday, 9 September 2010 18:15:20 UTC