EP0831456A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
Speech recognition method and apparatus therefor
Assignee: Canon Inc
•
Filing: 1997-09-18
•
Granted: —
•
Expires: 2017-09-18
Abstract
Abstract The present invention aims to provide a high-speed speech recognition method of a high recognition rate, utilizing speaker models. For this purpose, the method of this invention executes an acoustic process on the input speech, then calculates a coarse output probability utilizing an unspecified speaker model, and calculates a fine output probability utilizing an unspecified speaker model and clustered speaker models, for the states estimated, by the result of coarse calculation, to contribute to the result of recognition. Candidates of recognition are extracted by a common language search based on the obtained result, and a fine language search is conducted on thus extracted candidates to determine the result of recognition.
Inventors
- 👤 Yasuhiro Komori
- 👤 Tetsuo Kosaka
- 👤 Masayuki Yamada
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-02 13:00
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/14; G10L15/142; G10L15/18; G10L15/183; G10L15/19
Claims Analysis
Total Claims: 34
Claim 1
Independent Claim
72 words
A speech recognition method comprising steps
of: entering a speech; calculating a coarse output probability of said
input speech, utilizing an unspecified speaker model; determining a fine output probability of said
input speech, in both of an unspecified speaker model
selected according to said coarse output probability
and plural speaker models clustered into plural speaker
classes; and determining the result of recognition of said
input speech, based on said fine output probability.
Claim 2
Dependent Claim
40 words
A speech recognition method according to claim
1, wherein said speaker classes are prepared in a
hierarchic structure in such a manner that each speaker
belongs, in a specified hierarchic level, to plural
speaker classes or to a speaker class.
Claim 3
Dependent Claim
59 words
A speech recognition method according to claim
1, wherein said speaker classes are prepared by steps
of: preparing models respectively corresponding to
sound properties from the input speech; determining a speaker class to which said input
speech belongs, according to the distance of said
prepared models; and memorizing the model of said input speech as said
determined speaker class.
Claim 4
Dependent Claim
20 words
A speech recognition method according to claim
3, wherein said model prepared for each sound property
is a phoneme HMM.
Claim 5
Dependent Claim
57 words
A speech recognition method according to claim
3, wherein said models for respective sound properties
are phoneme models, and the sum of the distances
between the models, in the mutually corresponding
states of the phoneme models used for calculating said
distance of the models, is used as the distance between
the models for determining said speaker class.
Claim 6
Dependent Claim
20 words
A speech recognition method according to claim
3, wherein a Bhattacharyya distance is used as said
distance between the models.
Claim 7
Dependent Claim
21 words
A speech recognition method according to claim
3, wherein a Kullback information amount is used as
said distance between the models.
Claim 8
Dependent Claim
20 words
A speech recognition method according to claim
3, wherein an Euclid distance is used as said distance
between the models.
Claim 9
Dependent Claim
50 words
A speech recognition method according to claim
1, wherein: a language search is conducted according to the
result of calculation of said output probability and a
judgment according to predetermined language
limitation; and the result of said language search is outputted as
the result of recognition of said input speech.
Claim 10
Dependent Claim
82 words
A speech recognition method according to
claim 9, wherein, in said output probability
calculation and language search: the output probability calculation of a portion to
be calculated in common is conducted by a forward
search; and the result of recognition of said input speech is
determined by executing the output probability
calculation in portions other than said portion to be
calculated in common by a backward search, for each of
the unspecified speaker model and the models of the
clustered speaker classes.
Claim 11
Dependent Claim
50 words
A speech recognition method according to
claim 9, wherein, in the language search for said
plural speaker class models; determining a predetermined output probability
based on the output probability determined at each time
of plural speaker classes; and a common language search is executed, based on
said predetermined output probability.
Claim 12
Dependent Claim
23 words
A speech recognition method according to
claim 11, wherein said predetermined output probability
is taken as the output probability of the unspecified
speaker.
Claim 13
Dependent Claim
27 words
A speech recognition method according to
claim 11, wherein said predetermined output probability
is taken as the maximum value in the output
probabilities of the speaker classes.
Claim 14
Dependent Claim
24 words
A speech recognition method according to
claim 1, wherein said unspecified speaker model is
taken as an upper hierarchic model of said speaker
models.
Claim 15
Dependent Claim
23 words
A speech recognition method according to
claim 1, wherein a fine re-calculation of said output
probability is executed for all the speaker models.
Claim 16
Dependent Claim
43 words
A speech recognition method according to
claim 1, wherein the calculation of result of analysis
of said input speech and of output probability of the
unspecified speaker model is executed, utilizing scaler
quantization and dimensionally independent high-speed
output probability calculation method for HMM.
Claim 17
Dependent Claim
98 words
A speech recognition apparatus comprising: input means for entering a speech; coarse output probability calculation means for
calculating a coarse output probability of said input
speech, utilizing an unspecified speaker model; fine output probability calculation means for
determining a fine output probability of said input
speech, in both of an unspecified speaker model
selected according to said coarse output probability
determined by said coarse output probability
calculation means and plural speaker models clustered
into plural speaker classes; and recognition result determination means for
determining the result of recognition of said input
speech, based on said fine output probability.
Claim 18
Dependent Claim
40 words
A speech recognition apparatus according to
claim 17, wherein said speaker classes are prepared in
a hierarchic structure in such a manner that each
speaker belongs, in a specified hierarchic level, to
plural speaker classes or to a speaker class.
Claim 19
Dependent Claim
87 words
A speech recognition apparatus according to
claim 17, further comprising: sound property model preparation means for
preparing models respectively corresponding to sound
properties from the input speech; speaker class determination means for determining
a speaker class to which said input speech belongs,
according to the distance of said prepared models
prepared by said sound property model preparation
means; and speaker model preparation means for memorizing the
model of said input speech as the speaker class
determined by said speaker class determination means,
thereby preparing said speaker model.
Claim 20
Dependent Claim
20 words
A speech recognition apparatus according to
claim 19, wherein said model prepared for each sound
property is a phoneme HMM.
Claim 21
Dependent Claim
57 words
A speech recognition apparatus according to
claim 19, wherein said models for respective sound
properties are phoneme models, and the sum of the
distances between the models, in the mutually
corresponding states of the phoneme models used for
calculating said distance of the models, is used as the
distance between the models for determining said
speaker class.
Claim 22
Dependent Claim
20 words
A speech recognition apparatus according to
claim 19, wherein a Bhattacharyya distance is used as
said distance between the models.
Claim 23
Dependent Claim
21 words
A speech recognition apparatus according to
claim 19, wherein a Kullback information amount is used
as said distance between the models.
Claim 24
Dependent Claim
20 words
A speech recognition apparatus according to
claim 19, wherein an Euclid distance is used as said
distance between the models.
Claim 25
Dependent Claim
62 words
A speech recognition apparatus according to
claim 17, further comprising: language search means for conducting a language
search according to the result of calculation of said
output probability and a judgment according to
predetermined language limitation; wherein said recognition result determination
means is adapted to determine the result of said
language search means as the result of recognition of
said input speech.
Claim 26
Dependent Claim
70 words
A speech recognition apparatus according to
claim 25, wherein said language search means is adapted
to execute the output probability calculation of a
portion to be calculated in common by a forward search,
and to execute the output probability calculation in
portions other than said portion to be calculated in
common by a backward search, for each of the
unspecified speaker model and the models of the
clustered speaker classes.
Claim 27
Dependent Claim
57 words
A speech recognition apparatus according to
claim 25, wherein said language search means is
adapted, in the language search for said plural speaker
class models, to determine a predetermined output
probability based on the output probability determined
at each time of plural speaker classes, and to execute
a common language search, based on said predetermined
output probability.
Claim 28
Dependent Claim
23 words
A speech recognition apparatus according to
claim 27, wherein said predetermined output probability
is taken as the output probability of the unspecified
speaker.
Claim 29
Dependent Claim
27 words
A speech recognition apparatus according to
claim 27, wherein said predetermined output probability
is taken as the maximum value in the output
probabilities of the speaker classes.
Claim 30
Dependent Claim
24 words
A speech recognition apparatus according to
claim 17, wherein said unspecified speaker model is
taken as an upper hierarchic model of said speaker
models.
Claim 31
Dependent Claim
23 words
A speech recognition apparatus according to
claim 17, wherein a fine re-calculation of said output
probability is executed for all the speaker models.
Claim 32
Dependent Claim
43 words
A speech recognition apparatus according to
claim 17, wherein the calculation of result of analysis
of said input speech and of output probability of the
unspecified speaker model is executed, utilizing scaler
quantization and dimensionally independent high-speed
output probability calculation method for HMM.
Claim 33
Dependent Claim
87 words
A speech recognition method comprising the
steps of: receiving input speech signals to be processed; performing a rough recognition processing of the
received speech signals, using a generalized speaker
model; performing a refined recognition processing of the
received speech signals, using plural speaker models
which are clustered into plural speaker classes and using
an unspecified speaker model selected in accordance with
the results of the rough recognition processing; and determining a recognition result for the received
speech signals based on the outputs of said refined
recognition processing.
Claim 34
Dependent Claim
72 words
A speech recognition method comprising the
steps of: performing a rough recognition processing on
received speech signals, using a generalized speaker
model; selecting plural speaker models which are clustered
into plural speaker classes in dependance upon the
results of said rough recognition processing; performing a refined recognition processing using
the selected speaker models; and determining a recognition result for said received
speech signals, based on the results of said refined
recognition processing.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.