Managing agent engagement in a man-machine dialog
Agents engage and disengage with users intelligently. Users can tell agents to remain engaged without requiring a wakeword. Engaged states can support modal dialogs and barge-in. Users can cause disengagement explicitly. Disengagement can be conditional based on timeout, change of user, or environmental conditions. Engagement can be one-time or recurrent. Recurrent states can be attentive or locked. Locked states can be unconditional or conditional, including being reserved to support user continuity. User continuity can be tested by matching parameters or tracking user by many modalities including microphone arrays, cameras, and other sensors.
1. A method of controlling an engagement state of an agent during a dialog with a user, wherein the agent has dormant, initial and locked states, the method comprising:
determining whether the agent is in the dormant state, the initial state or the locked state;
when the determining determines that the agent is in the dormant state:
responsive to detecting a wake-up indicator, entering the initial state;
when the determining determines that the agent is in the initial state:
receiving an initial spoken request, from a reference user, that is a locking request that specifies a locking condition that specifies a voice class, wherein the specified voice class identifies a particular class of voice and encompasses a plurality of voices;
processing the initial spoken request; and
entering the locked state; and
when the determining determines that the agent is in the locked state and while the agent is in the locked state:
receiving a new spoken request from a new user;
applying a voice classifier to the new spoken request from the new user to produce a voice class hypothesis of the new spoken request;
determining that the locking condition is satisfied when the voice class hypothesis of the new spoken request matches the voice class specified by the locking condition;
in response to the locking condition being satisfied, processing the new spoken request and recognizing the new spoken request as an explicit unlock request; and
responsive to recognizing the unlock request, transitioning to the dormant state.
2. The method of claim 1 , further comprising:
when the determining determines that the agent is in the initial state:
building a reference characterization of the reference user; and
storing the reference characterization; and
when the determining determines that the agent is in a reserved state and while the agent is in the reserved state:
building a new characterization of the new user;
comparing the new characterization to the reference characterization to identify whether there is a match between the new characterization and the reference characterization; and
responsive to the comparing identifying the match, performing the processing of the new spoken request.
3. The method of claim 2 , wherein characteristics of the reference user are adjusted after identifying the match.
4. The method of claim 2 , wherein at least one of the reference characterization and the new characterization includes a voiceprint.
5. The method of claim 2 , wherein at least one of the reference characterization and the new characterization includes one or more scene analysis features extracted from one or more sensory signals.
6. The method of claim 5 , wherein the scene analysis features include an azimuth measurement.
7. The method of claim 5 , wherein the scene analysis features include a distance measurement.
8. The method of claim 5 , wherein the scene analysis features support face recognition.
9. The method of claim 2 , further comprising:
adjusting the reference characterization by repeatedly performing these steps:
building a snapshot characterization of the reference user;
comparing the snapshot characterization to the reference characterization to identify whether there is a mismatch between the snapshot characterization and the reference characterization;
responsive to finding the mismatch, entering the dormant state; and
adjusting the reference characterization based on the snapshot characterization.