RE: Requirement for UA / SS protocol

Barge-in is a feature that most end-users of speech will be familiar with.  To implement in Speech HTML, the UA would start both a synthesis and recognition session simultaneously.  If speech is detected before synthesis completes, then some entity needs to cancel/mute the incoming audio.  If synthesis completes before speech detected, timers need to be enabled to detect a potential lack or surplus of speech.


If we disallow the UA sending mid-request events to the SS, then either:

  * The synthesis and recognition requests must be handled by the same service and potentially in the same request.  This seems restrictive.

  * Need to implement this within the UA.  The main problem here is dealing with defunct requests.  Do we allow the synthesis to continue to stream even for very long requests after the UA has muted?  How do we inform the recognizer that timeout has occurred, and it should not longer listen?




Solving this natively within HTTP seems difficult.  I'm not sure, but I think it would boil down to impolitely tearing down the connection.  Bi-directional events may be difficult, but perhaps worth the effort.  Two ideas:

  * Use client chunking to send events and/or speech data to server.  Use "100 Continue" responses to send events and/or speech data from server.  This is probably a diversion from the original intention of Continue, but it doesn't seem to violate the spec.

  * Use HTTP as a simple request/response socket, and build a session one layer above.  This "session" would be unnecessary/optional for simple requests, but provide a structure to more complicated interactions.  For example, this could be used to perform channel adaptation, dialog continuity, advanced multi-modal interactions, bargein, etc.


Thoughts?



-----Original Message-----
From: Bjorn Bringert [mailto:bringert@google.com] 
Sent: Tuesday, November 23, 2010 9:03 AM
To: Young, Milan
Cc: Robert Brown; public-xg-htmlspeech@w3.org
Subject: Re: Requirement for UA / SS protocol

Yeah, that's a general problem with using HTTP chunking for
bidirectional streaming. There is some work going on in IETF to
address that I believe. I guess web app -> server events would not be
impossible if the HTTP streaming issue could be resolved.

About re-recognition, that would add a lot of complexity to both the
web app API and the server API since there would need to be a way to
refer to previous recognitions, and resource management to make sure
that servers don't have to hang on to all audio indefinitely. I would
to prefer to keep this out of v1 of our APIs.

About priorities, I think that the web app API is definitely most
important, since it is needed for both default and app-specified
speech services, and because it's the one that is most clearly part of
the HTML Speech XG charter. Typically, W3C does not specify
client-server protocols at all. Adding a protocol for network speech
services is a stretch of our charter, and the only chance to complete
it in our timeline would be to keep it super simple.

/Bjorn

On Tue, Nov 23, 2010 at 4:46 PM, Young, Milan <Milan.Young@nuance.com> wrote:
> I think I was missing something :).  Although chunking would encode the client events just fine, the server would probably have to delay it's response until the client has completed.
>
>
>
> -----Original Message-----
> From: Young, Milan
> Sent: Tuesday, November 23, 2010 8:32 AM
> To: 'Bjorn Bringert'; Robert Brown
> Cc: public-xg-htmlspeech@w3.org
> Subject: RE: Requirement for UA / SS protocol
>
> Hello Bjorn,
>
> Perhaps I'm missing something, but sending events from the client to server isn't any more difficult than sending events from server to client.  I suspect both could be symmetrically implemented in HTTP 1.1 through chunking.
>
> Regarding priorities, I believe the protocol is just as important as the web API.  As long as we have the concrete protocol proposal in hand along a similar timeline as the concrete web API proposals, we should be safe.
>
> Thanks
>
>
> -----Original Message-----
> From: Bjorn Bringert [mailto:bringert@google.com]
> Sent: Tuesday, November 23, 2010 4:23 AM
> To: Robert Brown
> Cc: Young, Milan; public-xg-htmlspeech@w3.org
> Subject: Re: Requirement for UA / SS protocol
>
> While I agree that those two requirements have use cases, I think that
> they are not important enough to warrant to increased complexity that
> they would add to any user agent - speech recognizer protocol. I can't
> see any simple way to implement them e.g. using HTTP 1.1.
>
> Since the XG has a very aggressive timeline, I think that the first
> priority should be to come up with a well-specified web app - user
> agent API. The next priority would be to add a basic user agent -
> speech service protocol that can handle everything required by the web
> app API. I'm quite sure that we will not have time to work on any
> advanced speech service API.
>
> /Bjorn
>
> On Mon, Nov 22, 2010 at 11:59 PM, Robert Brown
> <Robert.Brown@microsoft.com> wrote:
>> Hi Milan,
>>
>> * App-to-server events:
>>
>> Okay, consider my arm twisted :-)  Nice example.
>>
>> * Re-recognition:
>>
>> If we make this a requirement of the protocol, we're going to have to be careful about the design.  If the UA doesn't buffer the audio for re-recognition, this means a server somewhere is holding state, which may create an opportunity for evil apps/UAs to do bad things to services.  I guess this supports your point about increased effort speccing the protocol.
>>
>> -----Original Message-----
>> From: Young, Milan [mailto:Milan.Young@nuance.com]
>> Sent: Friday, November 19, 2010 8:00 PM
>> To: Robert Brown; Bjorn Bringert
>> Cc: public-xg-htmlspeech@w3.org
>> Subject: RE: Requirement for UA / SS protocol
>>
>> Hello Robert,
>>
>>
>> Let's say the user is booking a flight.  They are presented with a multi-modal UI in which they can type or speak any of the component facts (date, city, etc).
>>
>> If the implementation had to start and stop the recognition every time a new piece of data was filled in on the visual side, you'd run the risk of clipping.  It would be better if the visual could feed into the recognition stream with events (e.g. City=Seattle) in just the same way the recognition engine could feed events back to the visual.
>>
>> Keep in mind that although we haven't seen this class of advanced visual/speech integrations outside the lab, they are most certainly coming.  I believe this Speech HTML is going to unleash a flurry of innovation.
>>
>>
>>
>> On the re-recognition front, I do not support the UI buffering and resending the data.  This is a needless use of UA memory, bandwidth, and entails inherent latencies.  The only significant downside I can think of is increased effort specing the protocol.  But we have to do something similar for text interpretation anyway, so I don't think it's such a big deal.
>>
>>
>>
>> Thank you
>>
>>
>> -----Original Message-----
>> From: Robert Brown [mailto:Robert.Brown@microsoft.com]
>> Sent: Friday, November 19, 2010 4:25 PM
>> To: Young, Milan; Bjorn Bringert
>> Cc: public-xg-htmlspeech@w3.org
>> Subject: RE: Requirement for UA / SS protocol
>>
>> * App-to-server events:
>>
>> I think this example illustrates the concept, but I'm skeptical that that particular application would ever be built.  If I can scroll through the list, it's probably a short list, so I doubt the recognizer will need any hints.  And I may as well just touch the person's name, since I can see it right there on the screen.  It feels academic to me.
>>
>> * Re-recognition:
>>
>> Another re-recognition example is the user saying "Starbucks in Redmond Washington".  The initial grammar doesn't have all the business names for all the districts in the entire continent, but it is good at detecting which district the user is talking about.  The re-recognition is then done using a business listing grammar that's specific to that district.
>>
>> But I think there may be an argument that re-recognition is not a hard requirement of the protocol.  The UA could just as easily buffer the audio, and re-send it for a second recognition if required by the app.
>>
>> Maybe there should just be a requirement: "The Web Application should be able to re-use the audio recording from one recognition to perform another recognition."
>>
>> -----Original Message-----
>> From: Young, Milan [mailto:Milan.Young@nuance.com]
>> Sent: Friday, November 19, 2010 8:20 AM
>> To: Bjorn Bringert; Robert Brown
>> Cc: public-xg-htmlspeech@w3.org
>> Subject: RE: Requirement for UA / SS protocol
>>
>> First the use cases:
>>
>> Web-app to SS events - The user wants to place a call, but can't quite remember the name of the contact.  The visual UI scrolls through the list of contacts and sends an event each time a new contact is displayed.  The recognizer uses this information to weight the recognition result because it's likely the user will speak a name it has just seen.
>>
>> Re-recognition using previous audio - The user has requested directory assistance to find a residential phone number.  They thought the contact lived in City-A, but no relevant results found.  They want to try again in City-B, and shouldn't have to be asked again for the utterance.
>>
>> Interpretation over text - A common approach in speech processing is to use a large statistical or speaker-dependent model to identify the lexical nature of the tokens.  A second pass over the data extracts meaning from the tokens.  At present, this second pass seems to be a more difficult task, and sometimes several attempts need to be made each with a different base context.  For example the word "bill" might refer to a financial transaction, a duck, a hat, or a person.
>>
>>
>>
>> Other responses:
>>
>> FPR11 - I'd like to add a second sentence to this statement (either in summary or text).  "This includes both TBD standard and extension parameters."
>>
>> FPR28 and 29 - Perhaps these statements could be slightly adjusted with "... fire implementation-specific events to the web app".
>>
>> FPR4 - EMMA seems like a sufficiently flexible standard to handle all foreseeable needs.  Perhaps now is not the time, but I'd like to see if we can agree to make this a required part of the protocol.
>>
>>
>> Thanks
>>
>>
>>
>> -----Original Message-----
>> From: Bjorn Bringert [mailto:bringert@google.com]
>> Sent: Friday, November 19, 2010 3:17 AM
>> To: Robert Brown
>> Cc: Young, Milan; public-xg-htmlspeech@w3.org
>> Subject: Re: Requirement for UA / SS protocol
>>
>> We already had a requirement that there must be a standard protocol.
>> If I understand this list correctly, it adds a number of requirements on what features this standard protocol must support. I propose that we consider each of the bullet points a separate requirement, so that they can be discussed independently.
>>
>> I think that most of them look fine. The only two that I'm not sure about are:
>>
>> - web-app -> speech service events, with the same objection that Robert raised.
>>
>> - Re-recognition using previous audio streams. What's the use case for this?
>>
>>
>> Also, I think that the following are already covered by existing requirements:
>>
>> - "Both standard and extension parameters passed from the web app to the speech service at the start of the interaction.  List of standard parameters TBD."
>>  Covered by "FPR11. If the web apps specify speech services, it should be possible to specify parameters."
>>
>> - The speech service -> web app part of the birirectional events requirement is covered by:
>>
>> FPR21. The web app should be notified that capture starts.
>> FPR22. The web app should be notified that speech is considered to have started for the purposes of recognition.
>> FPR23. The web app should be notified that speech is considered to have ended for the purposes of recognition.
>> FPR24. The web app should be notified when recognition results are available.
>> FPR28. Speech recognition implementations should be allowed to fire implementation specific events.
>> FPR29. Speech synthesis implementations should be allowed to fire implementation specific events.
>>
>> - "EMMA results passed from the SS to the web app.  The syntax of this result is TBD (e.g. XML and/or JSON)."
>> Covered by:
>> FPR4. It should be possible for the web application to get the recognition results in a standard format such as EMMA.
>>
>> - "Interpretation over text."
>> Covered by (if I understand it correctly):
>> FPR2. Implementations must support the XML format of SRGS and must support SISR.
>>
>>
>> So, the remaining requirements from Milan's list that I support adding are:
>>
>> * At least one standard audio codec.  UAs are permitted to advertise alternate codecs at the start of the interaction and SSs are allowed to select any such alternate (e.g. HTTP Accept).
>>
>> * Transport layer security (e.g. HTTPS) if requested by the web app.
>>
>> * Session identifier that could be used to form continuity across multiple interactions (e.g. HTTP cookies).
>>
>> /Bjorn
>>
>> On Fri, Nov 19, 2010 at 1:49 AM, Robert Brown <Robert.Brown@microsoft.com> wrote:
>>> I mostly agree.  But do we need bidirectional events?  I suspect all
>>> the interesting ones originate at the server: start-of-speech;
>>> hypothesis; partial result; warnings of noise, crosstalk, etc.  I'm
>>> trying to think why the server would care about events from the
>>> client, other than when the client is done sending audio (which it
>>> could do in response to a click or end-point detection).
>>>
>>>
>>>
>>> From: public-xg-htmlspeech-request@w3.org
>>> [mailto:public-xg-htmlspeech-request@w3.org] On Behalf Of Young, Milan
>>> Sent: Thursday, November 18, 2010 5:34 PM
>>> To: public-xg-htmlspeech@w3.org
>>> Subject: Requirement for UA / SS protocol
>>>
>>>
>>>
>>> Hello,
>>>
>>>
>>>
>>> On the Nov 18th conference, I volunteer to send out proposed wording
>>> for a new requirement:
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> Summary - User agents and speech services are required to support at
>>> least one common protocol.
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> Description - A common protocol will be defined as part of the final
>>> recommendation.  It will be built upon some TBD existing application
>>> layer protocol and include support for the following:
>>>
>>>
>>>
>>>   * Streaming audio data (e.g. HTTP 1.1 chunking).  This include both
>>> audio streamed from UA to SS during recognition and audio streamed
>>> from SS to UA during synthesis.
>>>
>>>
>>>
>>>   * Bidirectional events which can occur anytime during the interaction.
>>> These events could originate either within the web app (e.g. click) or
>>> the SS (e.g. start-of-speech or mark) and must be transmitted through
>>> the UA in a timely fashion.  The set of events include both standard
>>> events defined by the final recommendation and extension events.
>>>
>>>
>>>
>>>   * Both standard and extension parameters passed from the web app to
>>> the speech service at the start of the interaction.  List of standard
>>> parameters TBD.
>>>
>>>
>>>
>>>   * EMMA results passed from the SS to the web app.  The syntax of
>>> this result is TBD (e.g. XML and/or JSON).
>>>
>>>
>>>
>>>   * At least one standard audio codec.  UAs are permitted to advertise
>>> alternate codecs at the start of the interaction and SSs are allowed
>>> to select any such alternate (e.g. HTTP Accept).
>>>
>>>
>>>
>>>   * Transport layer security (e.g. HTTPS) if requested by the web app.
>>>
>>>
>>>
>>>   * Session identifier that could be used to form continuity across
>>> multiple interactions (e.g. HTTP cookies).
>>>
>>>
>>>
>>>   * Interpretation over text.
>>>
>>>
>>>
>>>   * Re-recognition using previous audio streams.
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> Thank you
>>>
>>>
>>
>>
>>
>> --
>> Bjorn Bringert
>> Google UK Limited, Registered Office: Belgrave House, 76 Buckingham Palace Road, London, SW1W 9TQ Registered in England Number: 3977902
>>
>>
>>
>
>
>
> --
> Bjorn Bringert
> Google UK Limited, Registered Office: Belgrave House, 76 Buckingham
> Palace Road, London, SW1W 9TQ
> Registered in England Number: 3977902
>



-- 
Bjorn Bringert
Google UK Limited, Registered Office: Belgrave House, 76 Buckingham
Palace Road, London, SW1W 9TQ
Registered in England Number: 3977902

Received on Wednesday, 1 December 2010 20:50:20 UTC