Signal conditioned minimum error rate training for continuous speech recognition

• US 5,806,029 A
• Filed: 09/15/1995
• Issued: 09/08/1998
• Est. Priority Date: 09/15/1995
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Buhrke, Eric Rolfe, Chou, Wu, Rahim, Mazin G.
• Primary Examiner(s): Knepper, David D.
• Application Number:
  US08/528,821
• Time in Patent Office: 1,089 Days
• Field of Search: 395/2.09, 395/2.35, 395/2.36, 395/2.42, 395/2.52, 395/2.54, 704/200, 704/226, 704/227, 704/233, 704/243-245
• US Class Current: 704/244
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/144 Training of HMMs
• CPC Class Codes: G10L 21/0272 Voice signal separating

Abstract

Hierarchical signal bias removal (HSBR) signal conditioning uses a codebook constructed from the set of recognition models and is updated as the recognition models are modified during recognition model training. As a result, HSBR signal conditioning and recognition model training are based on the same set of recognition model parameters, which provides significant reduction in recognition error rate for the speech recognition system.

Claims

1. A method of signal conditioning for removing an unknown signal bias in a speech signal in a speech recognition system storing a set of recognition models, comprising the following steps:

(A) generating a feature signal which characterizes features of the speech signal, the feature signal comprising one or more frames of feature vectors;

(B) storing the feature signal in memory;

(C) constructing a codebook comprising one or more clusters based on the set of recognition models;

(D) calculating a cluster-specific bias for each of the clusters of the codebook;

(E) calculating a cluster-specific weight for each of the clusters of the codebook;

(F) generating a frame-dependent weighted bias signal for each frame of the feature signal;

(G) subtracting the frame-dependent weighted bias signal for each frame of the feature signal from each frame of the feature signal to generate a conditioned feature signal; and

(H) storing the conditioned feature signal in memory to replace the feature signal.

2. A method according to claim 1, wherein step (A) comprises the steps:

receiving the speech signal; and

extracting features from the speech signal to generate the feature signal.

3. A method according to claim 1, further comprising the step:

(I) doubling the number of clusters within the codebook.

4. A method according to claim 3, further comprising the step:

repeating steps (A) through (I) a preselected number of times.

5. A method according to claim 4, further comprising the step:

determining whether the number of clusters is equal to a preselected number.

6. A method according to claim 1, further comprising the step:

using the conditioned feature signal to modify the set of recognition models.

7. A method according to claim 1, wherein:

the codebook is a condensed representation of the set of recognition models.