Recognizing speech by selectively canceling model function mixture components

• US 6,999,929 B2
• Filed: 09/05/2001
• Issued: 02/14/2006
• Est. Priority Date: 09/06/2000
• Status: Expired due to Fees

• Current Assignee: Sony Deutschland GmbH (Sony Group Corp.)
• Original Assignee: Sony Limited (Sony Group Corp.)
• Inventors: Kompe, Ralf, Goronzy, Silke
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US09/947,109
• Publication Number:
  US 20020046024A1
• Time in Patent Office: 1,623 Days
• Field of Search: 704/255, 704/266, 704/256, 704/236, 704/231, 704/234, 704/233, 704/251, 704/242, 704/270, 704/270.1, 704/235, 381/43, 382/224
• US Class Current: 704/255
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

A method for recognizing speech is proposed wherein the process of recognition is started using the starting acoustic model (SAM) and wherein the current acoustic model (CAM) is modified by removing or cancelling model function mixture components (MFM_jk) which are negligible for the description of the speaking behavior and quality of the current speaker. Therefore, the size of the acoustic model (SAM, CAM) is reduced by adaptation to the current speaker enabling fast performance and increased recognition efficiency.

Claims

1. Method for recognizing speech,

wherein, for the process of recognition, a current acoustic model (CAM) based on a set of model function mixtures (MFM1, . . . , MFMn) is used, and

wherein said current acoustic model (CAM) is adapted during the recognition process by changing at least in part the contributions of model function mixture components (MFMjk) of model function mixtures (MFMj) based on at least one recognition result already obtained,

characterized in that

the process of recognition is started using a starting acoustic model (SAM) as said current acoustic model (CAM),

after a given number of performed recognition steps and/or after a given number of obtained recognition results, a modified acoustic model (MAM) is generated based on said current acoustic model (CAM) by canceling model function mixture components (MFMjk) having negligible contributions with respect to at least a given number of recognition results already obtained, and

the process of recognition is continued using said modified acoustic model (MAM) as said current acoustic model (CAM) in each case.

2. Method according to claim 1,

wherein a modified acoustic model (MAM) is generated repeatedly after each fixed and/or predetermined number of performed recognition steps and/or obtained recognition results, in particular after each single performed recognition step and/or obtained recognition result.

3. Method according to claim 1,

wherein the number of recognition steps and/or recognition results after which a modified acoustic model (MAM) is generated is determined and/or changed within the current process of recognition and/or adaptation.

4. Method according to claim 1,

wherein an acoustic model is used—

in particular as said starting acoustic model (SAM) and/or as said current acoustic model (CAM) in each case—

the model function mixtures (MFMj) of which at least contain distribution functions or the like, in particular functions of the Gaussian type or the like and/or in particular as said model function mixture components (MFMjk).

5. Method according to claim 1,

wherein each of said model function mixtures (MFMj) is based on a function vector (f_j) and a weight factor vector (a_j), each of which having a finite and/or equal number (n_j) of components.

6. Method according to claim 5,

wherein each of said model function mixtures (MFMj) is a linear combination or superposition of its vector function components (f_j,k) weighted by its weight factor vector components (a_j,k), particularly represented by a scalar product of the weight factor vector (a_j) and the function vector (f_j):

$\mathrm{MFMj}=\underset{k=1}{\overset{n_j}{\sum <br><br>}}<br><br>a_{j,k}<br><br>f_{j,k}=a_j^T<br><br>f_j=a_j·<br><br>f_j$

where MFMj denotes the j^th model function mixture, a_j denotes a j^th weight factor vector with a_j,k being its k^th component, f_j denotes the j^th function vector with f_j,k being its k^th component, a_j^T denotes the transposed form of a_j and denotes the scalar or inner product of the vectors.

7. Method according to claim 5,

wherein each of said model function mixture components (MFMjk) is classified as being negligible if the absolute value (|a_j,k|) of its weight factor vector component (a_j,k) is beyond a given threshold value (c_j,k), in particular for a given number (m_j,k) of times of recognition steps already performed and/or recognition results already obtained.

8. Method according to claim 7,

wherein each of said threshold values (c_j,k) is predetermined and/or fixed, in particular for each of the model function mixture components (MFMjk) independently and in particular before starting the recognition process.

9. Method according to claim 7,

wherein each of said threshold value (c_j,k) is determined and/or modified during the recognition process, in particular based on signal quality information of the speech input and/or in particular with respect to statistical and/or noise data.

10. Method according to claim 1,

wherein weight factor vector components (a_j,k) are modified among other components of the modified acoustic model (MAM) for speaker adaptation, in particular to reduce certain weight factor vector components (a_j,k) below certain thresholds.