Watson Dispatch Manager
Call for Code 2018
Team 1, Dublin, Ireland
Solution Outline
 Proposed solution augments Emergency Service Agent (ESA) by
 1--Converting Emergency call's audio stream to a transcript ready for use by ESA
 2--Identify all entities ( nouns of relevance ) in the transcript for populating any fields in
Emergency Dispatch System
 3--Search and provide closely matching emergency calls of recent past with similar entity values
for consideration as duplicates or supplementary information
 4--In case of waiting queue at Emergency Response number, engage user and collect
information for action by ESA.
 Note: Not all of the above is targeted for 31st August, 2018 milestone.
Use Cases and User research
 There are two possible use cases for this solution
 Functionality while an ESA is engaged with a caller (Passive case)
 Functionality to interact with caller while no ESA is available to engage. (Active case)
 Examples of why Active case would be relevant
 Example from latest natural disaster: Floods in Kerala India - https://indianexpress.com/article/india/kerala-floods-
rains-rescue-inside-kochis-sos-centre-5312671/
 https://www.sundaypost.com/fp/999-callers-on-hold-for-20-minutes-as-pressure-on-nhs-rises999-emergencylives-at-
risk-as-desperate-callers-are-put-on-hold-for-20-minutesnhs-winter-crisis-ae-wards-under-siege-as-virulent-flu/
 https://globalnews.ca/news/2376143/911-on-hold-how-other-provinces-answer-emergency-calls/
 The team engaged with Emergency dispatchers in Ireland’s dispatch service (999) to understand the value
proposition and the outcome is that target ESA population would be interested in such a solution.
Architecture of Solution
IBM CloudApplication Layer
MH
Orchestration
Application
Call Platform
Text To Speech
Emergency Caller
Web UI
Watson Assistant
WatsonOrchestration
Db Storage
HTTPS
(JSON)
ESA
.WAV /
.AIFF
Speech To Text
Natural Language Understanding
Watson Discovery Service
Phase 2
Information Flow Phase 1
 For active use case, Agent platform invokes Watson Assistant service using api when all ESA are busy.
 Watson Assistant (WA) initiates conversation and sends initiation message from WA to (Text to Speech)
TTS service and relays audio output to Agent platform.
 Audio input is captured from the Agent platform using broadband sampling and passed on as .wav /
.aiff packet to ‘Speech To Text’ which then provides transcript output.
 Transcript output is used by Watson Assistant to identify intent and entities associated with the incident.
 WA provides closing message to caller (through TTS) with summary of entity values and user interface
shows a match with an existing incident.
User Interface Phase 1
Entity values
Location on
map
Transcription
Information Flow Phase 2
 For active use case, Agent platform invokes Watson Orchestrator using its api when all ESA are busy.
 WO invokes Watson Assistant (WA) to initiate conversation and sends initiation message from WA to
TTS service and relays audio output to Agent platform.
 Audio input is captured from the Agent platform using broadband sampling and passed on as .wav /
.aiff packet to WO which then relays the output to ‘Speech To Text’ service and obtains transcript
output.
 WO calls Natural Language Understanding service to obtain vital entity values.
 WO provides closing message to caller (through TTS) with summary of entity values.
 WO updates the Watson Discovery Service collection with the latest incident and queries for similar
reported incidents in last ‘n’ minutes using entity values in last callers narration.
 WO compares search result set to query parameters and presents matches and differences on the web
UI for the ESA to review and action.
 WO pushes enriched data from WDS collection to db storage.
 For passive use case, Watson Assistant would not be involved.
Phase 1 Goals
 Transcription of Audio
 Basic conversational interaction (welcome and closure)
 Limited entity detection
 Basic comparison
 UI with Transcription, Entity and comparison display
Refinement and possibilities (Phase 2 post 31/08)
 Improvement in transcription accuracy with language model and acoustic model.
 Conversational improvement to ask for missing entities and disambiguation
 Improvement in entity detection and relationship detection through Watson Knowledge
Studio
 Improvement in User Interface and comparison of incidents.
 Analytics in UI example: based on location entity by interface to google maps.

Watson Dispatch Manager

  • 1.
    Watson Dispatch Manager Callfor Code 2018 Team 1, Dublin, Ireland
  • 2.
    Solution Outline  Proposedsolution augments Emergency Service Agent (ESA) by  1--Converting Emergency call's audio stream to a transcript ready for use by ESA  2--Identify all entities ( nouns of relevance ) in the transcript for populating any fields in Emergency Dispatch System  3--Search and provide closely matching emergency calls of recent past with similar entity values for consideration as duplicates or supplementary information  4--In case of waiting queue at Emergency Response number, engage user and collect information for action by ESA.  Note: Not all of the above is targeted for 31st August, 2018 milestone.
  • 3.
    Use Cases andUser research  There are two possible use cases for this solution  Functionality while an ESA is engaged with a caller (Passive case)  Functionality to interact with caller while no ESA is available to engage. (Active case)  Examples of why Active case would be relevant  Example from latest natural disaster: Floods in Kerala India - https://indianexpress.com/article/india/kerala-floods- rains-rescue-inside-kochis-sos-centre-5312671/  https://www.sundaypost.com/fp/999-callers-on-hold-for-20-minutes-as-pressure-on-nhs-rises999-emergencylives-at- risk-as-desperate-callers-are-put-on-hold-for-20-minutesnhs-winter-crisis-ae-wards-under-siege-as-virulent-flu/  https://globalnews.ca/news/2376143/911-on-hold-how-other-provinces-answer-emergency-calls/  The team engaged with Emergency dispatchers in Ireland’s dispatch service (999) to understand the value proposition and the outcome is that target ESA population would be interested in such a solution.
  • 4.
    Architecture of Solution IBMCloudApplication Layer MH Orchestration Application Call Platform Text To Speech Emergency Caller Web UI Watson Assistant WatsonOrchestration Db Storage HTTPS (JSON) ESA .WAV / .AIFF Speech To Text Natural Language Understanding Watson Discovery Service Phase 2
  • 5.
    Information Flow Phase1  For active use case, Agent platform invokes Watson Assistant service using api when all ESA are busy.  Watson Assistant (WA) initiates conversation and sends initiation message from WA to (Text to Speech) TTS service and relays audio output to Agent platform.  Audio input is captured from the Agent platform using broadband sampling and passed on as .wav / .aiff packet to ‘Speech To Text’ which then provides transcript output.  Transcript output is used by Watson Assistant to identify intent and entities associated with the incident.  WA provides closing message to caller (through TTS) with summary of entity values and user interface shows a match with an existing incident.
  • 6.
    User Interface Phase1 Entity values Location on map Transcription
  • 7.
    Information Flow Phase2  For active use case, Agent platform invokes Watson Orchestrator using its api when all ESA are busy.  WO invokes Watson Assistant (WA) to initiate conversation and sends initiation message from WA to TTS service and relays audio output to Agent platform.  Audio input is captured from the Agent platform using broadband sampling and passed on as .wav / .aiff packet to WO which then relays the output to ‘Speech To Text’ service and obtains transcript output.  WO calls Natural Language Understanding service to obtain vital entity values.  WO provides closing message to caller (through TTS) with summary of entity values.  WO updates the Watson Discovery Service collection with the latest incident and queries for similar reported incidents in last ‘n’ minutes using entity values in last callers narration.  WO compares search result set to query parameters and presents matches and differences on the web UI for the ESA to review and action.  WO pushes enriched data from WDS collection to db storage.  For passive use case, Watson Assistant would not be involved.
  • 8.
    Phase 1 Goals Transcription of Audio  Basic conversational interaction (welcome and closure)  Limited entity detection  Basic comparison  UI with Transcription, Entity and comparison display
  • 9.
    Refinement and possibilities(Phase 2 post 31/08)  Improvement in transcription accuracy with language model and acoustic model.  Conversational improvement to ask for missing entities and disambiguation  Improvement in entity detection and relationship detection through Watson Knowledge Studio  Improvement in User Interface and comparison of incidents.  Analytics in UI example: based on location entity by interface to google maps.