EP0903727A1
Expired / Lapsed
EP
⏱️ 0.0 years remaining
A system and method for automatic speech recognition
Assignee: Istituto Trentino di Cultura
•
Filing: 1997-09-17
•
Granted: —
•
Expires: 2017-09-17
Abstract
Abstract Word sequences comprised in a speech signal are recognized by applying a language model that is the result of a combination of other language models and capable of being dynamically redefined during the man-machine interaction. Use is made of hierarchical language models (HLMs) comprising appropriate terminal symbols capable of constituting references to other language models comprised in a plurality of language models. The said hierarchical language models (HLMs) are represented in the form of probabilistic recurrent transition networks with the use of a search engine of the recursive type that operates directly on the said networks and is capable of performing in a single step both the recognition action and the parsing function of an incoming utterance.
Inventors
- 👤 Fabio Brugnara
- 👤 Marcello Federico
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-09 12:24
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/18; G10L15/183; G10L15/19; G10L15/193; G10L15/197; G10L15/22; G10L2015/226; G10L2015/228
Claims Analysis
Total Claims: 9
Claim 1
Independent Claim
106 words
An automatic speech signal recognition process, wherein
the speech word sequences are recognized by applying (3) a
language model derived from a combination of a plurality of
language models, the said model being capable of being
redefined at every new input phrase on the basis of predefined
internal rules, characterized in that the said language models
(LMs) are hierarchical language models (HLMs) and comprise
respective terminal symbols capable of constituting references
to at least another language model of the said plurality and
in that the recognition of the said speech word sequences is
translated into appropriate actions to modify a corresponding
physical entity (10 to 14).
Claim 2
Dependent Claim
48 words
A process in accordance with Claim 1, characterized in
that the said hierarchical language models (HLMs) are
represented as probabilistic recurrent transition networks
whose graphs comprise arcs with appropriate labels capable of
constituting references to other networks that represent
respective other hierarchical language models of the said
plurality.
Claim 3
Dependent Claim
57 words
A process in accordance with Claim 2, characterized in
that the said hierarchical language models (HLMs) are defined
in a dynamic manner by combining the elements of a set of
predefined basic language models by compilation thereof during
the performance of the recognition action in the said
probabilistic recurrent transition networks, maintaining
possibly references to other networks.
Claim 4
Dependent Claim
49 words
A process in accordance with Claim 2 or Claim 3,
characterized in that it comprises the operation of
implementing a direct recursive search function in the said
probabilistic recurrent transition networks that realizes in a
single step (3) both the recognition and the parsing function
of an incoming utterance.
Claim 5
Dependent Claim
140 words
An automatic speech recognition system comprising means
(2, 3, 6, 7) for recognizing speech word sequences that apply
(3) a language model deriving from a combination (8) of a
plurality (5) of language models, the said model being capable
of being redefined (7) for every new input (1) on the basis of
predefined internal rules, characterized in that it comprises: decoding means (3, 6) acting on the speech signal, where
the said decoding means (3) utilize hierarchical language
models (HLMs) comprising terminal symbols capable of
constituting references to other language models of the said
plurality of language models (LMs), and operating means (7) that cooperate with the said
decoding means (3, 6) to translate the recognition of the said
speech word sequences into appropriate actions to modify a
corresponding physical entity (10 to 14) and to update the
language model.
Claim 6
Dependent Claim
60 words
A system in accordance with Claim 5, characterized in
that in the said decoding means (3, 6) the said hierarchical
language models are represented in the form of probabilistic
recurrent transition networks (Figure 3, Figure 4) whose
graphs comprise arcs with appropriate labels capable of
constituting references to the respective networks of other
hierarchical language models of the said plurality.
Claim 7
Dependent Claim
52 words
A system in accordance with Claim 6, characterized in
that the said operating means (7, 8) generate the said
hierarchical language models in dynamic form by combining the
elements of a set of basic language models stored in the
system (4, 5) and compiling them in the appropriate
probabilistic recurrent transition networks.
Claim 8
Dependent Claim
49 words
A system in accordance with Claim 6 or Claim 7,
characterized in that it comprises a recursive search engine
(3) that operates directly on the said probabilistic recurrent
transition networks to realize simultaneously and in a single
step both the recognition and the parsing operation of the
incoming utterances.
Claim 9
Independent Claim
100 words
A system in accordance with any of Claims 5 to 8,
characterized in that the said operating means (7) realize at
least one of the operations chosen from among the following
group: transcription of the said speech signal into electronic
text, automatic bringing into action of an apparatus upon
receipt of the said speech signal, entry of structured data
corresponding to word sequences of the said speech signal into
a database, extraction of information in response to a request
corresponding to the word sequences of the said speech signal,
or automatic generation of vocal response/request messages for
a system user.
References Cited (Backward Citations)
EP0533338A2
Cited
WO1994016434A1
Cited
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.