Speaker independent speech recognition method utilizing multiple training iterations

• US 5,806,034 A
• Filed: 08/02/1995
• Issued: 09/08/1998
• Est. Priority Date: 08/02/1995
• Status: Expired due to Term

• Current Assignee: Exelis Incorporated (L3Harris Technologies, Inc.)
• Original Assignee: ITT Corporation (ITT, Inc.)
• Inventors: Huang, William Y., Bahler, Lawrence G., Naylor, Joe A.
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Storm, Donald L.
• Application Number:
  US08/510,321
• Time in Patent Office: 1,133 Days
• Field of Search: 395/2.65, 395/2.49, 395/2.5, 395/2.52, 395/2.45, 395/2.48, 395/2.63, 395/2.64, 395/2.43, 395/2.51, 395/2.47, 395/2.53, 395/2.54
• US Class Current: 704/256
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

A method for recognizing spoken utterances of a speaker is disclosed, the method comprising the steps of providing a database of labeled speech data; providing a prototype of a Hidden Markov Model (HMM) definition to define the characteristics of the HMM; and parameterizing speech utterances according to one of linear prediction parameters or Mel-scale filter bank parameters. The method further includes selecting a frame period for accommodating the parameters and generating HMMs and decoding to specified speech utterances by causing the user to utter predefined training speech utterances for each HMM. The method then statistically computes the generated HMMs with the prototype HMM to provide a set of fully trained HMMs for each utterance indicative of the speaker. The trained HMMs are used for recognizing a speaker by computing Laplacian distances via distance table lookup for utterances of the speaker during the selected frame period; and iteratively decoding node transitions corresponding to the spoken utterances during the selected frame period to determine which predefined utterance is present.

Claims

1. A method for recognizing spoken utterances of a speaker, comprising the steps of:

providing a database of labeled speech data;

providing a prototype of a Hidden Markov Model (HMM) definition to define the characteristics of the HMM;

parameterizing speech utterances according to one of linear prediction parameters or Mel-scale filter bank parameters;

selecting a frame period of approximately 20 msecs and a window duration of substantially 25 msecs for accommodating said parameters;

generating HMMs and decoding to specified speech utterances by causing said speaker to utter predefined training speech utterances for each said HMM,

statistically computing said generated HMMs with said prototype HMM to provide a set of fully trained HMMs for each utterance indicative of said speaker;

using said trained HMMs for recognizing a speaker by computing Laplacian distances for utterances of said speaker during said selected frame period; and

iteratively decoding node transitions corresponding to said spoken utterances during said selected frame period to determine which predefined utterance is present.

2. The method according to claim 1, further including the step of hand marking said speech data for seeding the training of HMMs said step of hand marking occurring between the steps of parameterizing said speech data and selecting a frame period for accommodating said parameters.

3. The method according to claim 1, wherein said node transitions correspond to node transition paths extending through a lattice, wherein said node transition paths are identical at each iteration.

4. The method according to claim 3, wherein said node decoding is viterbi decoding.

5. The method according to claim 3, wherein the step of decoding further includes providing a backpointer vector for optimizing optimal paths to each of the nodes.

6. The method according to claim 1, wherein the step of generating HMMs further includes a preemphasis coefficient of approximately 1.0 and a scale log energy setting of approximately 1.0.

7. The method according to claim 6, wherein said fully trained HMMs are trained in at least 40 training passes.

8. The method according to claim 6, wherein said fully trained HMMs are trained in approximately 100 training passes.

9. The method according to claim 1 wherein said Laplacian distances are stored in a lookup table.