[asterisk-app-dev] WebSocket Stasis Control Best Practice

[Matthew Jordan]

On Wed, Jun 18, 2014 at 12:16 PM, Ben Klang <bklang at mojolingo.com> wrote:
> Excuse my somewhat tardy reply to this thread, but since you brought up AMD:
>
> On Jun 16, 2014, at 11:47 AM, Ben Langfeld <ben at langfeld.me> wrote:
>
> On Sun, Jun 15, 2014 at 9:24 PM, Krandon <krandon.bruse at gmail.com> wrote:
>>
>> Hello Asterisk friends,
>>
>> I am currently interfacing with Asterisk through ARI and loving the
>> experience so far. I have successfully originated calls and dumped them into
>> my Stasis app. I am trying to figure out what the best way is to send a
>> channel into an Application. The current architecture for
>> /channels/{id}/play works well for the majority of my app, but I am running
>> into a block figuring out how to interact with Asterisk dialplan
>> applications.
>>
>> To give an example - I submit an originate to go to SIP/vendor/phoneNumber
>> - with the other leg going to App: myStasisApp, {"soundFile":"blah"}. That
>> works fine (with the proper quote escaping). Now my Stasis app has received
>> the channelID to which we can do a lot of neat stuff. Say I play a sound to
>> the user but then want to call the app WaitForSilence. What's the best way
>> to do this? I may be misinterpreting the intended use of both Stasis and ARI
>> - but I am curious to see what your thoughts are.
>>
>> Also, for the stasis app to get a list of arguments, I am passing it
>> through as JSON. So far that is working fine - but I wanted to see if there
>> was a better way to get a list/array of app args to Stasis.
>>
>> Forgive me if there is an easy solution - through digging and poking the
>> last few days, I have not been able to find the intended use case or even a
>> use case.
>>
>>
>
> Well, the solution for this just got added into the Asterisk 12 branch, and
> so it hasn't made it into a release yet. It should be coming soon in
> Asterisk 12.4.0.
>
> The TALK_DETECT [1] function enables AMI/ARI events [2] [3] [4] [5] on a
> channel, such that a connected ARI application receives notifications over
> the WebSocket when a person starts/stops talking. This lets you
> asynchronously 'know' when both talking/silence has occurred - obviating the
> need for the WaitForSilence/WaitForNoise dialplan applications. Plus,
> because it is asynchronous, if you decide you don't *want* to wait for
> silence, you don't have to!
>
> With a bit of manipulation, you could also construct AMD from this as well -
> but I'll admit that's a bit more challenging. I'd be interested in people's
> experiences with attempting to do that, and if an asynchronous "IS_HUMAN"
> detection function is needed or not.
>
>
> We are in the process right now of creating an application that needs
> asynchronous AMD.  Specifically, we are implementing LumenVox’s CPA
> product[1] and the use case is this:
>
> * Reminder call is placed to recipient
> * Recipient answers (don’t yet know if it is a human or a machine)
> * Outgoing message begins to play
> * If a human is detected, stop playback and connect to an agent
> * If a machine is detected, keep playing back until…
> * If a beep is detected, stop and restart playback
>
> The only way to achieve this is if we can have an async speech recognizer
> running while simultaneously playing output, which isn’t possible with
> Dialplan today, and would require a specialized app even if it were
> implemented that way.  Instead, we are hoping to have a lower-level
> primitive to do signals detection and playback asynchronously.

You are correct that dialplan constructs (such as SpeechBackground) do
not readily translate to asynchronous handling of events. While there
are limited situations where 'event handlers' have crept into the
dialplan - such as pre-dial/hangup handlers - those are (a) special
cases and (b) not readily applicable to a large swath of events in
Asterisk.

Dialplan/AGI being synchronous was a large motivating factor for ARI.
While synchronous interfaces are easier to understand, we don't live
in a synchronous world.

> In an ideal world, ARI would provide primitives for playback (file or TTS)
> and input (DTMF or ASR). Some more background from discussion related to our
> project, courtesy Ben Langfeld:
>
> The asynchronous example is more complex. While Adhearsion sees both the
> input and output components as being asynchronous, this is a fake facility
> provided by Punchblock to make Asterisk look like an async server when it is
> not. Both components are implemented atop synchronous Asterisk dialplan
> applications:
>
> For output: Playback() or MRCPSynth()
> For input: MRCPRecog()
>
> This means that given the simplest approach to implementation discussed
> above, the output would be executed, followed by the input being queued and
> executed once the output had completed. If we were to swap the two, not only
> would we now have a coordination problem where we have to queue cancellation
> of the output to paper over the race condition introduced by potentially
> being asked to stop it before we have a handle on it, we would have the same
> blocking problem with MRCPRecog().
>
> So that rules out combining one of the UniMRCP dialplan applications with
> the Playback() application in this fashion. There are two other remaining
> solutions that come to mind:

Agreed; the existing mechanisms in the dialplan do not translate well
to asynchronous control of a channel.

Without punting all the way to ARI, in Asterisk 12 you do also have
the option of asynchronously stopping media during a Playback
operation via the AMI action ControlPlayback [1]. Granted, that
doesn't solve many of the synchronous issues you are describing, but
it would at least let you cancel a Playback operation.

[1] https://wiki.asterisk.org/wiki/display/AST/Asterisk+12+ManagerAction_ControlPlayback

> A prompt command to combine the output and input into a single dialplan
> application invocation (MRCPRecog() for native file playback,
> SynthAndRecog() for TTS). This avoids the problem of multiple dialplan
> applications blocking one another, but introduces a fresh one: these
> applications terminate output as soon as recognition completes (or earlier
> if barge-in is enabled). There is no opportunity to inject logic to filter
> the recognition result prior to terminating the output, nor do I think this
> would make sense.
>
> The Asterisk Speech API (SpeechLoadGrammar(), SpeechActivateGrammar(),
> SpeechStart(),SpeechBackground(), etc). If SpeechBackground() this would be
> the obvious solution, but it unfortunately is not. SpeechBackground()
> actually sits in a loop, directing audio frames to the recognizer while
> simultaneously rendering frames of audio (the first option is a file path).
> The app does not return until recognition has completed, so cannot be
> combined with Playback(). Upon recognition completion, the output will be
> terminated, regardless of the recognition result, so this suffers the same
> problem as Rayo Prompt. It is also not possible to use any other output
> renderer, such as a TTS engine via MRCP.
>
> Can we implement Asterisk/Lumenvox CPA in way to be compatible with the
> adhearsion-cpa controller methods API?
>
> The problems stated above leave us with only one option: extra capability
> must be introduced to Asterisk in order to handle simultaneous dialplan
> applications, or to introduce a true async version of SpeechBackground().
> The viability of this is something that must be discussed with the Asterisk
> project / Digium. Note that FreeSWITCH already has this capability, but
> would also need less invasive changes to cope with LumenVox CPA as stated
> above; a far more approachable task.

A few thoughts here:
(1) I'm not sure that introducing a dialplan variant of
SpeechBackground that had some asynchronous capabilities will buy
much. At the end of the day, you're still stuck in the dialplan -
which has a synchronous model of operation. To do everything that you
need, you need:
  (a) Asynchronous results from the speech engine
  (b) Asynchronous capabilities to control media operations
  (c) Asynchronous capabilities to control the speech recognition
While (b) does exist in the previously mentioned AMI action, we're now
once again requiring a combination of AGI/dialplan + AMI - which is
clunky. It's the reason why we wrote ARI in the first place!
(2) The good news is, the speech API in Asterisk is not synchronous.
The current APIs that expose it certainly are, but there is no
implicit long running blocking operation involved with
ast_speech_write (or any of the other C API functions involved in
res_speech). Building an asynchronous function that emits events
(similar to TALK_DETECT) or adding this as an explicit operation to an
ARI resource is not a very hard task. In fact, using audiohooks is a
fairly painless way of passing audio frames from a channel (regardless
of where they are) into ast_speech_write, and would be a simple way of
passing media into the speech engine in an asynchronous fashion.
(3) I think it'd be nice if this was a native operation in ARI. Unlike
TALK_DETECT - which is a relatively simple on/off use case - there's a
lot of subtlety to speech recognition. Some of the existing operations
(such as engine creation/enabling) could probably be hidden under an
operation on a channel resource, but the ability to activate certain
grammars while speech recognition is enabled on a channel would
certainly be nice. I'd imagine this would be somewhat similar to the
/play operation, where what you are handed back is a resource that has
some additional properties that can be manipulated independently.
Something like:

POST /channels/{id}/recognizeSpeech?speechId=12345&default_grammar=yes_no

A speech resource (maybe a different name? We typically use a plural
form for this - speechInstances?) could be used to manipulate an
active speech recognition process on a channel:

DELETE /speech/12345/  (stop speech recognition)

POST /speech/12345/grammar?name=moar_grammars

POST /speech/12345/parameter?name=engine_specific_property&value=foobar

Or other things along those lines.


As an aside, this is a very subjective statement that feels
inappropriate given the context of this discussion:

{quote}
"Note that FreeSWITCH already has this capability, but would also need
less invasive changes to cope with LumenVox CPA as stated above; a far
more approachable task."
{quote}

Much of the core of this is analogous to TALK_DETECT - which I wrote I
think in two evenings on my couch while watching reruns of Law & Order
SVU. I wouldn't say this is an extremely difficult task. It may be a
bit more than a few 'couch evenings', but we're not talking about
overhauling the bridging framework here. I don't know FreeSWITCH from
Adam, so I won't comment on how difficult it is to implement there.
However, I'd prefer to keep comments about the relative implementation
difficulty of something within Asterisk versus FreeSWITCH out of these
discussions in the future.

> In short, the adhearsion-cpa API can be honoured for the synchronous
> detection case trivially. It cannot be honoured for the async case, nor can
> any equivalent alternative be introduced, without changes to Asterisk.
>

Agreed - this will definitely require changes to Asterisk, and the
most appropriate place to implement this functionality is probably
ARI.

Matt

-- 
Matthew Jordan
Digium, Inc. | Engineering Manager
445 Jan Davis Drive NW - Huntsville, AL 35806 - USA
Check us out at: http://digium.com & http://asterisk.org