Speaker adaptation device and speech recognition device

• US 6,223,159 B1
• Filed: 12/22/1998
• Issued: 04/24/2001
• Est. Priority Date: 02/25/1998
• Status: Expired due to Fees

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Ishii, Jun
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US09/217,928
• Time in Patent Office: 854 Days
• Field of Search: 704/251, 704/231, 704/256, 704/255, 704/250, 704/200, 704/201, 704/257, 704/240, 704/247, 704/252
• US Class Current: 704/256.7
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

Voice feature quantity extractor extracts feature vector time-series data by acoustic feature quantity analysis of the speaker'"'"'s voice. Reference speaker-dependent conversion factor computation device computes reference speaker-dependent conversion factors through use of a reference speaker voice data feature vector and an initial standard pattern. The reference speaker-dependent conversion factors are stored in a reference speaker-dependent conversion factor storage device. Speaker-dependent conversion factor selector selects one or more sets of reference speaker-dependent conversion factors stored in the reference speaker-dependent conversion factor storage device. Speaker-dependent conversion factor computation device computes speaker-dependent conversion factors, through use of the selected one or more sets of reference speaker-dependent conversion factors. Speaker-dependent standard pattern computation device converts parameters of the initial standard pattern, through use of the speaker-dependent conversion factors, and thus-converted parameters are output as a speaker-dependent standard pattern.

Claims

1. A speaker adaptation device comprising:

a voice feature quantity extractor which extracts feature vector time-series data by acoustic feature quantity analysis of a speaker'"'"'s voice;

a reference speaker-dependent conversion factor computation device which computes reference speaker-dependent conversion factors through use of a reference speaker voice data feature vector and an initial standard pattern;

a reference speaker-dependent conversion factor storage device which stores the reference speaker-dependent conversion factors computed by the reference speaker-dependent conversion factor computation device;

a speaker-dependent conversion factor selector which selects one or more sets of reference speaker-dependent conversion factors from the reference speaker-dependent conversion factors stored in the reference speaker-dependent conversion factor storage device, with regard to the feature vector time-series data extracted by the voice feature quantity extractor;

a speaker-dependent conversion factor computation device which computes speaker-dependent conversion factors through use of the selected one or more sets of the reference speaker-dependent conversion factors; and

speaker-dependent standard pattern computation device which converts parameters of the initial standard pattern through use of the speaker-dependent conversion factors computed by the speaker-dependent conversion factor computation device and which outputs a speaker-dependent standard pattern.

2. The speaker adaptation device as defined in claim 1, wherein the reference speaker-dependent conversion factor computation device computes, with regard to the reference speaker voice data feature vector, reference speaker-dependent conversion factors for each reference speaker, each reference speaker group including a plurality of reference speakers, or each reference speaker as well as each reference speaker group including a plurality of reference speakers;

and wherein the computed reference speaker-dependent conversion factors are stored in the reference speaker-dependent conversion factor storage device.

3. The speaker adaptation device as defined in claim 1, wherein the reference speaker-dependent conversion factor computation device computes reference speaker-dependent conversion factors for each parameter group divided into a plurality of segments with regard to the initial standard pattern, or each reference speaker group including reference speakers with regard to the reference speaker data feature vector as well as each parameter group divided into a plurality of segments with regard to the initial standard pattern;

and wherein the computed reference speaker-dependent conversion factors are stored in the reference speaker-dependent conversion factor storage device.

4. The speaker adaptation device as defined in claim 1, wherein the reference speaker-dependent conversion factor computation device computes reference speaker-dependent conversion factors based on a computation using a conversion matrix and a constant vector, both of which depend on each reference speaker, based on a computation using only a conversion constant vector depending on each reference speaker, or based on a combined computation of a conversion matrix and a constant vector, both of which depend on each reference speaker, and a conversion constant vector depending on each reference speaker;

and wherein the computed reference speaker-dependent conversion factors are stored in the reference speaker-dependent conversion factor storage device.

5. The speaker adaptation device as defined in claim 1, wherein the speaker-dependent conversion factor selector selects one or more reference speaker-dependent conversion factors having conversion factor similarity by computation of a similarity between the speaker-dependent conversion factors which are produced through use of the feature vector time-series data extracted by the voice feature quantity extractor and the initial standard pattern and the reference speaker-dependent conversion factor stored in the reference speaker-dependent conversion factor storage device, one or more reference speaker standard pattern conversion factors in order of likelihood from a largest to a lowest by computation of likelihood of the feature vector time-series data extracted by the voice feature quantity extractor with regard to the reference speaker-dependent standard pattern produced through use of the reference speaker-dependent conversion factor stored in the reference speaker-dependent conversion factor storage device and the initial standard pattern, or one or more reference speaker standard pattern conversion factors having conversion factors similarity from a largest conversion factor similarity to a lowest conversion factor similarity, and likelihood;

and wherein the speaker-dependent conversion factor computation device computes speaker-dependent conversion factors through use of the selected one or more reference speaker standard pattern conversion factors.

6. The speaker adaptation device as defined in claim 1, wherein the speaker-dependent conversion factor computation device subjects to weighted averaging operations the one or more sets of reference speaker-dependent conversion factors selected by the speaker-dependent conversion factor selector, thereby computing one or more sets of speaker-dependent conversion factors.

7. The speaker adaptation device as defined in claim 1, wherein a continuous mixture distribution hidden Markov model is used as the initial standard pattern.

8. The speaker adaptation device as defined in claim 7, wherein the reference speaker-dependent conversion factor computation device computes a reference speaker-dependent conversion factor which converts a mean vector, by means of an element distribution function constituting a symbol output probability density function of the continuous mixture distribution hidden Markov model used as the initial standard pattern, comprising a Gaussian distribution function or a Laplace distribution function.

9. The speaker adaptation device of claim 1, wherein said speaker adaptation device is Part of a speech recognition device.

10. A speaker adaptation device comprising:

a reference speaker-dependent conversion factor storage unit that stores reference speaker-dependent conversion factors;

a speaker-dependent conversion factor computation device, which computes speaker-dependent conversion factors using select reference speaker-dependent conversion factors from said reference speaker-dependent conversion factor storage unit;

a speaker-dependent standard pattern computation device which computes a speaker-dependent standard pattern based on said speaker-dependent conversion factors computed by said speaker-dependent conversion factor computation device and an initial standard pattern.