Speaker Identification and Verification

1
Speaker Identification and Verification

• Dan Burnett, Nuance
• 58th IETF

2
Terminology

• Speaker identification -- using utterances from a
  speaker, determine who the caller is out of a set
  of known speakers
• Speaker verification -- using utterances from a
  speaker, determine whether the caller is who
  he/she claims to be (requires an identity claim)
• Training -- using utterances from a speaker to
  train a unique voiceprint that can later be used
  to identify/verify a speaker. Applies to both
  SI/SV.

3
draft-burnett-mrcpext-00.txt

• Created by Nuance and Intervoice
• Proposes extensions to MRCP v1 (draft-shanmugham-
  mrcp-04.txt)
• Based originally on Nuance functionality,
  modified to be more general
• Starting point for MRCP v2 functionality
  discussions
• Also extensions for speaker-enrolled grammars,
  hotword recognition, and to the recognition
  resource

4
Proposed SI/SV process(simplified, see section
6.7)
VER-START-SESSION
VER-BUFFERING-START
VER-SET-VOICEPRINT
VER-BUFFERING-CONTROL
VER-FROM-BUFFER
VER-END-SESSION
VER-BUFFERING-STOP
Requires active buffering and ver/id sessions.
5
Discussion points

• Why buffering?
• Registry for return info
• Anything else before I convert to MRCPv2?

6
Voice/Text Grammar Enrollment(simplified, see
section 5.5)
START-ENROLLMENT-SESSION

• Extension to existing recognition resource
• Creates speaker-produced grammar entries
• E.g., voice-enrolled entries for voice dialing
• Both speech and text can be used to create
  grammar entries

PAUSE/RESUME-ENROLLMENT-SESSION
ENROLLMENT-ROLLBACK
RECOGNIZE/STOP
ADD/DELETE/MODIFY-PHRASE
END/ABORT-ENROLLMENT-SESSION
These methods already exist in the recognizer
resource
7
Hotword(see section 7)

• New recognition resource
• Instead of listening for a set time period,
  listens continuously until it matches a grammar
• Non-matching speech is ignored and does not
  affect the state of the recognizer

8
Other Extensions

• Record method (sec. 4.4)
• Allows end-pointed recording of an audio stream
• Interpret method (sec. 4.5)
• Behaves as a recognition except that text input
  is given instead of an audio stream. It returns
  a standard recognition result minus any
  audio-specific values.