Speech recognition with language-dependent model vectors

• US 7,630,878 B2
• Filed: 05/04/2004
• Issued: 12/08/2009
• Est. Priority Date: 07/28/2003
• Status: Active Grant

• Current Assignee: Cerence Operating Company (Cerence Inc.)
• Original Assignee: SVOX AG (Microsoft Corporation)
• Inventors: Stan, Sorel, Fingscheidt, Tim
• Primary Examiner(s): Lerner; Martin
• Application Number:
  US10/566,293
• Publication Number:
  US 20070112568A1
• Time in Patent Office: 2,044 Days
• Field of Search: 704/8, 704/231, 704/233, 704/243, 704/255, 704/256, 704/277, 379/88.06
• US Class Current: 704/8
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2015/223 Execution procedure of a spoken command

Abstract

Speaker-dependent speech recognition is performed upon detecting a speech signal encompassing a voice command. The speech signal is divided into time frames and characterized in each detected time frame by forming a corresponding property vector. A language-independent feature vector sequence is formed from one or several property vectors and then stored. The language-independent feature vector sequence is allocated to a language-dependent sequence of model vectors in a speech resource having a plurality of model vectors. A piece of allocation information indicating allocation of the language-independent feature vector sequence to a language-dependent sequence of model vectors is stored, then the voice command allocated to the model vector sequence is identified.

Claims

1. A method for speaker-dependent speech recognition, comprising:

capturing a speech signal, including a speech command, of a speaker;

breaking down the speech signal into time frames;

characterizing the speech signal in each captured time frame by forming a corresponding feature vector;

forming a language-independent feature vector sequence from at least one feature vector;

storing the language-independent feature vector sequence;

assigning the language-independent feature vector sequence to a language-dependent sequence of model vectors in a first language resource which includes a multiplicity of language-dependent model vectors;

storing first assignment information which specifies assignment of the language-independent feature vector sequence to the language-dependent sequence of model vectors;

recognizing the speech command which is assigned to the language-dependent sequence of model vectors;

selecting a second language resource different from the first language resource;

assigning the language-independent feature vector sequence previously stored to a language-dependent model vector sequence in the second language resource; and

storing second assignment information regarding said assigning of the language-independent feature vector sequence to the language-dependent model vector sequence in the second language resource.

2. A method as claimed in claim 1, wherein the speech signal is made up of acoustic units.

3. A method as claimed in claim 2, wherein each of the first and second language resources is based on a Hidden Markov Modeling of acoustic units of a speech signal.

4. A method as claimed in claim 3, wherein an acoustic unit is formed by a word or a phoneme.

5. A method as claimed in claim 3, wherein an acoustic unit is formed by word segments or groups of related phonemes.

6. A method as claimed in claim 5, wherein different language resources are assigned to at least one of different languages and different language environments.

7. A method as claimed in claim 6, wherein different language environments indicate different environmental noise situations.

8. A method as claimed in claim 7, further comprising reducing dimensionality of the feature vector or the language-independent feature vector sequence by a matrix multiplication before assigning to the model vector or the model vector sequence.

9. A method as claimed in claim 8, further comprising specifying the matrix for dimensional reduction from one of a Linear Discriminant Analysis, a Principal Component Analysis and an Independent Component Analysis.

10. A method as claimed in claim 9, wherein the speaker-independent speech recognition is language-dependent.

11. A communication device, comprising:

a microphone recording a speech signal, including a speech command, of a speaker;

a processor processing the speech signal by breaking down the speech signal into time frames, characterizing the speech signal in each captured time frame by forming a corresponding feature vector and forming a language-independent feature vector sequence from at least one feature vector;

a storage unit storing the language-independent feature vector sequence obtained from the speech signal; and

a speech recognition entity, coupled to the microphone, configured for at least speaker-dependent speech recognition by assigning the language-independent feature vector sequence to a language-dependent sequence of model vectors in a first language resource which includes a multiplicity of language-dependent model vectors, storing first assignment information which specifies assignment of the language-independent feature vector sequence to the language-dependent sequence of model vectors, recognizing the speech command which is assigned to the language-dependent sequence of model vectors, selecting a second language resource different from the first language resource, assigning the language-independent feature vector sequence previously stored to a language-dependent model vector sequence in the second language resource, and storing second assignment information corresponding thereto.

12. A communication device as claimed in claim 11, wherein said speech recognition entity simultaneously uses speaker-dependent and speaker-independent vocabularies.