Recognition of spoken and spelled proper names

1
Recognition of spoken and spelled proper names
Author Michael Meyer, Hermann Hild
Reporter CHEN, TZAN HWEI
2
Outline

• Introduction
• Experiments
• Summary

3
Introduction

• The recognition of increasingly large sets of
  spoken names is difficult
• Very large recognition vocabularies contain many
  easily confused words or even homophones.
• In this paper, it compares the performance for
  proper name recognition when a name is spoken
  only, spelled only or both spoken and spelled.

4
Introduction (cont)

• In what contexts do people speak and spell names

Table 1 Three scenarios for speaking and
spelling a proper name
5
Experiments

• Speech data
• A database of about 2800 German last name spoken
  by 57 different speaker, according to scenario 2.
• Recorded with a close-talking microphone at a
  sampling rate of 16 kHz.
• The boundaries between all spoken and spelled
  names were to identified to conduct scenario 1

6
Experiments (cont)

• The pronunciation dictionary covers about half of
  the 2800 names of our speech data.
• The set of 1337 spoken and spelled names is used
  all the experiments described below.
• For experiments, we use a MS-TDNN as a
  specialized letter recognizer.
• And use the LVCSR of the JANUS system as a spoken
  name recognizer.

7
Experiments (cont)

• JANUS
• 60.0 names correct were achieved on the test set
  of the 1337 spoken last names.
• To recognize the spelled name with JANUS, 93.3
  correct names were achieved on the spelled names.
• MS-TDNN
• Achieved 96.5 correct spelled names on the test .

8
Experiments (cont)

• Small list
• We assume that the list of names to be recognized
  is small enough, so that every name can be
  explicitly represented in the dictionary.
• How can we combine the different information
  provided by the spoken and spelled names?

9
Experiments (cont)

• After all, the pronunciations of the spelled
  letters represent in approximation the sounds of
  the letters in the fluently spoken words.
• TOM versus T-O-M
• Exceptions are letters with unusual pronunciation
  and those letter combination which define their
  own pronunciation, such as
• Sch , ch
• In the following we will just combines the two
  representations on the basis of their acoustic
  scores only

10
Experiments (cont)

• Scenario 1

11
Experiments (cont)

• Scenario 2
• With 86.1 correct, the recognition on the entire
  utterance is worse than on the spelled apart
  alone.
• It is possible to adapt a similar approach as in
  scenario 1.
• The boundary of the first best hypothesis was
  used for the weighting of all hypotheses.
• Resulting in a recognition rate of 89.1
• To incorporate the MS-TDNN letter recognizer,
  resulting in 95.8 recognition rate.

12
Experiments (cont)
Figure 1 names correct for a weighted
combination of the N-best list of spoken and
spelled names (scenario 1 and 2)
13
Experiments (cont)

• Large lists
• If the number of names exceeds the recognizers
  maximum vocabulary size, a different approach has
  to be taken.
• A two-step approach is employed.
• A coarse recognition run is used to get a reduced
  list of name candidates.
• Then, these are processed in which all the
  previously described techniques for small word
  lists can be applied.

14
Experiments (cont)

• In the case of scenario 1, the list of candidates
  can be easily reduced if only the spelled names
  are considered in the first pass.
• For scenario 2, only phonemes and letters in
  JANUSs recognition vocabulary.

15
Experiments (cont)

• For scenario 2
• A special language model is employed.
• A list of the most similar names can be
  retrieved, and then used in another JANUS
  recognition run.
• The letter segments are then re-recognized with
  the MS-TDNN

16
Experiments (cont)
Table 3 Summary of results for the separated and
combined recognition of fluently spoken and
spelled last names
17
Summary

• By combining the N-best lists of both the spoken
  and spelled recognition, the overall performance
  can be improved.
• An input of either L or FL can be distinguish
  with almost 99 correct, resulting of 95.5 names
  correct without a priori knowing whether L or FL
  was spoken.

18
Caller Identification from Spelled-Out Personal
Data Over the Telephone
Reporter CHEN, TZAN HWEI
19
Outline

• Introduction
• The personal identification algorithm
• Tests and results
• Conclusions

20
Introduction

• The problem of automatically identifying the
  caller in a telephone conversation from the
  information spoken in the call is extremely
  difficult.
• The identification must take place despite rather
  substantial speech recognition errors that may be
  made by the machine.

21
Introduction (cont)

• We can find a solution to the problem if we make
  two assumptions.
• We assume that there is a database of records
  containing personal information about there the
  caller which can serve as a reference during the
  identification process.
• We ask our caller to spell the personal
  identifying items so that the spoken vocabulary
  is small and we can look for correlations with
  the items in the database.

22
The personal identification algorithm
Fig 1. the algorithm of personal identification
from spelled tokens
23
The personal identification algorithm (cont)

• Bayesian computation which starts with an
  estimate, for each record in , of the
  probability that record represents the identity
  of the caller.
• It uses the acquired information and updates
  each records probability that it corresponds to
  the current callers identity.
• The incremental computation is

24
The personal identification algorithm (cont)

• Bayesian update of probabilities

25
Tests and results

• The system was tested using a database of one
  million records that was constructed by using
  random combinations of 4,375 female first names,
  1,129 male first names, and 88,799 last names.
• The account numbers were generated so that the
  values for the last four digits of the number
  occurred with equal frequency throughout the
  database.
• The city, postal code, and phone number fields
  were generated to correspond the locations in the
  U.K.

26
Tests and results (cont)

• Our test involved identifying 300 different
  records in the database.
• If the system was unable to make an
  identification of the target record after asking
  the user for all of the information, the caller
  was asked to make a second attempt using the same
  information.
• If the system failed to produce a result after
  the second attempt, the call was terminated at
  that point.

27
Tests and results (cont)

• For each telephone call, the users were asked
  eight questions
• Enter your ID, using you telephone keypad,
  followed by the pound key. If you make a mistake
  press the star key.
• You entered (the value entered in (1)) if this is
  correct, press 1. If it is not, press2.
• Please say the first four letters of your last
  name.
• Please say the first four digits of your first
  name.

28
Tests and results (cont)

• Please say the last four digits of your card
  number.
• Could you please spell the city currently listed
  on your account?
• Please say you phone number.
• Please say the postal code currently listed on
  your account.

29
Tests and results (cont)
Fig 2. summary of results from 300 calls.
30
Tests and results (cont)
Fig 3. Rate of ASR character misrecognition by
filed
31
Tests and results (cont)
Fig. 4. Rate of misrecognition of field by ASR
(misrecognition at least one error made in
spelled filed value)
32
Tests and results (cont)
Fig 5. Average cumulative number of records
examined by system
33
Conclusions

• The method tolerates high misrecognition rates.
• The method can be used with off-the-shelf
  component it doesnt require specialized ASR.
• To allow the personal identification information
  to be spoken instead of spelled tokens.

34
Record Rv that will be verified
Select another T
Request T from caller
Collect subset near T
Rm Rv
Another T?
No
operator
Add subset to S
Yes
No
Risk(Rv) lt Risk(reject)
No
Update the risk for each Record in S
Reject !
Yes
Rm lt- min risk in S
Accept !