Speech recognition

• US 5,146,503 A
• Filed: 08/24/1988
• Issued: 09/08/1992
• Est. Priority Date: 08/28/1987
• Status: Expired due to Fees

• Current Assignee: British Telecommunications PLC (BT Group PLC)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Millar, Paul C., Cameron, Ian R.
• Primary Examiner(s): Shaw, Dale M.
• Assistant Examiner(s): Knepper, David D.
• Application Number:
  US07/235,416
• Time in Patent Office: 1,476 Days
• Field of Search: 364/513.5, 381/36-43
• US Class Current: 704/255
• CPC Class Codes: G10L 15/063 Training

Abstract

A speech recognition system comprises a recogniser (2) for receiving speech signals from users. The recogniser (2) compares each received word with templates of words stored in a reference template store (3) and flags each template that corresponds most closely to a received word. The flagged templates are stored in a template store (11). The recogniser (2) compares the speech pattern from a given user of a second utterance of a word for which a flagged template is already stored in the template store (11) with the templates stored in the reference template store (3) and with the flagged templates in the template store (11), so as to produce a second flagged template of that word. The second flagged templates are also stored in the template store (11). Sifting means (21) analyse a group of flagged templates of the same word, and produce therefrom a second, smaller group of templates of the word. These templates are stored in another template store (22).

Claims

1. A method of constructing templates for use in speech recognition, the method comprising the steps of:

analysing speech signals from a plurality of users to construct at least one template for each word input by each user, and

constructing from said templates a set of representative templates for each of said words,

wherein during said constructing step each user inputs all the words of a given list of words in a "forced" mode, and then inputs at least some of the words of the list in a "free" mode by answering a series of questions.

2. A method as in claim 1, wherein:

each word input in the "forced" mode is compared with all templates of the words in the list, these templates being stored in a first template store,

the analysing step identifying the template that corresponds most closely to each word spoken in the "forced" mode and storing such identified template in a second template store.

3. A method as in claim 2, wherein the analysing step compares each word spoken in the "free" mode with all the templates stored in the first and second template stores, and identifies and stores the template that corresponds most closely to each word spoken in the "free" mode in the second template store.

4. A method as in claim 3, wherein:

all the templates of a given number of users are stored in a third template store prior to the constructing step, and

the templates identified during the constructing step are stored in a fourth template store.

5. A method as in claim 4, wherein:

the constructing step includes the step of comparing each template in the third template store with all the templates of the same word, and with all the templates of any other word associated with said first-mentioned word in a predetermined manner, the constructing step being such that the templates of a given word stored in the fourth template store are representative of a maximum number of the templates of that word stored in the third template store and unrepresentative of any other template stored in the third template store.

6. Apparatus for generating a set of speech templates to be used in subsequent speech recognition processes, said apparatus comprising:

means for generating first speech templates in a "forced" mode where a human speaker is prompted directly to speak a given word;

means for generating second speech templates in a "free" mode where a human speaker is induced to speak the same given word but without direct request for same; and

means for combining said first and second speech templates into a common set of speech templates for subsequent use in speech recognition processes.

7. Method for generating a set of speech templates to be used in subsequent speech recognition processes, said method comprising the steps of:

generating first speech templates in a "forced" mode where a human speaker is prompted directly to speak a given word;

generating second speech templates in a "free" mode where a human speaker is induced to speak the same given word but without direct request for same; and

combining said first and second speech templates into a common set of speech templates for subsequent use in speech recognition processes.

8. Apparatus for constructing a set of speech-representative templates, said apparatus comprising:

prompt means for generating an output prompting a user to speak a word;

wherein the prompt means is operable to prompt the user;

(a) in a "forced" mode, by indicating directly the word the user is to speak, and;

(b) in a "free" mode, by indicating a context which is calculated to induce the user to speak that word;

input means for receiving a word input by a user;

means for producing a template from a thus input word; and

template store means for storing a thus-produced template, whereby the template store may store, for each word, templates corresponding to both "forced" and "free" modes.

9. Apparatus according to claim 8, in which the prompt means in a speech synthesizer, operates

in the "forced" mode to utter the word the user is to speak, and

in the "free" mode to utter a question to which the said word is calculated to form the answer.

10. Apparatus according to claim 8 or 9 for constructing a set of speaker-independent templates, the apparatus being arranged to receive input words from a plurality of users and store a corresponding plurality of templates, the apparatus further comprising:

means connected to said template store means, for selecting a representative subset of said plurality of templates and for storing said subset of templates.

11. Apparatus according to claim 8 or 9 further comprising:

speech recognition means for identifying input words, and for allocating to thus-identified input words, data indicating their identity.