Speech segmentation

• US 6,055,495 A
• Filed: 04/30/1997
• Issued: 04/25/2000
• Est. Priority Date: 06/07/1996
• Status: Expired due to Term

• Current Assignee: Hewlett-Packard Development Company, L.P. (HP Inc.)
• Original Assignee: Hewlett-Packard Company (HP Inc.)
• Inventors: Collins, Michael John, Tucker, Roger Cecil Ferry
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US08/846,612
• Time in Patent Office: 1,091 Days
• Field of Search: 704/210, 704/200, 704/208, 704/243, 704/248, 704/215
• US Class Current: 704/210
• CPC Class Codes: G06F 3/16 Sound input; Sound output speech processing G10L
• CPC Class Codes: G10L 2025/783 based on threshold decision
• CPC Class Codes: H04M 1/6505 storing speech in digital form
• CPC Class Codes: H04M 2203/301 Management of recordings
• CPC Class Codes: H04M 3/533 Voice mail systems
• CPC Class Codes: H04M 3/53333 Message receiving aspects

Abstract

The present invention relates to the management of voice data.

Voice messages left on a recipient'"'"'s answerphone or delivered via a voicemail system are a popular form of person-to-person communication. Such voice messages are quick to generate for the sender but are relatively difficult to review for the recipient; speech is slow to listen to and, unlike inherently visual forms of messages such as electronic mail or handwritten notes, cannot be quickly scanned for the relevant information. The present invention aims to make it easier for users to find relevant information in voice messages, and other kinds of voice record, such as recordings of meetings and recorded dictation.

According to the present invention we provide a method of speech segmentation comprising processing speech data so as to detect putative pauses and characterised by forming speech block boundaries at a selected subset of the pauses, said selection being based on a preselected target speech block length.

The invention may be applied in an application where speech is represented visually.

Claims

1. A method of speech segmentation comprising processing speech data so as to detect pauses and characterised by forming speech block boundaries at a selected subset of the pauses, wherein said subset of said pauses is selected so as to approximate each speech block to a preselected target speech block length.

2. A method according to claim 1 comprising:

devising an error metric which indicates how much a speech block length deviates from the target speech block length;

calculating the error value for the speech data in a file;

detecting the most significant pause in the speech data;

calculating the average error value for the speech blocks on either side of the speech block boundary at the pause and retaining the speech block boundary in the speech data if said average error value is less than the current error value;

iteratively detecting the next most significant pause, calculating the average error value for the speech blocks on either side of the speech block boundary at the next most significant pause and retaining the speech block boundary in the speech data if this further reduces the error value.

3. A method according to claim 2 wherein the error value is proportional to the square of the difference between the duration of a speech block and the target speech block length.

4. A method according to claim 2 wherein the error value is proportional to the square of the difference between the square roots of the duration of a speech block and the target speech block length.

5. A method according to claim 2 comprising the further steps of:

for each speech block boundary which is retained, comparing the average error value for the speech blocks either side of the boundary with the error value for a block comprising the sum of the two blocks,

deleting the speech block boundary that gives the greatest error reduction,

repeating these steps until no further error value reduction is obtained.

6. A method according to claim 2, wherein the significance of a pause is proportional to the duration of the pause.

7. A method according to claim 2, wherein the significance of a pause is inversely proportional to the energy of the pause.

8. A method according to claim 1 comprising using a speech/silence detector to identify pauses before processing the speech data.

9. A method according to claim 1 comprising searching for the lowest energy interval of predetermined size in a segment of speech data and treating that lowest energy interval as the most significant pause in the respective speech segment.

10. A method of speech segmentation comprising:

processing speech data so as to detect pauses;

forming speech block boundaries at a selected subset of the pauses, selection being based on a preselected target speech block length, said selection accomplished by dividing the total duration of the speech data in a file by the target speech block length to derive a desired pause number n, and detecting the n most significant pauses in that file and forming speech block boundaries at these n pauses in the speech data.

11. A method according to claim 10 wherein the significance of a pause is proportional to the duration of the pause.

12. A method according to claim 10 wherein the significance of a pause is inversely proportional to the energy of the pause.

13. A method according to claims 10 comprising searching for the lowest energy interval of predetermined size in a segment of speech data and treating that lowest energy interval as the most significant pause in the respective speech segment.

14. A system for implementing a method of speech segmentation comprising:

detector means for processing speech data so as to detect pauses; and

segmenter means for forming speech block boundaries at a selected subset of the pauses, said selected subset being selected so as to aproximate each speech block to a preselected target speech block length.

15. A system according to claim 14 further comprising display means for representing speech data in the form of speech blocks.

16. A system for implementing a method of speech segmentation comprising:

detector means for processing speech data so as to detect pauses; and

segmenter means for forming speech block boundaries at a selected subset of the pauses, selection of said selected subset being such as to approximate each speech block to a preselected target speech block length, said selection accomplished by dividing the total duration of the speech data in a file by the target speech block length to derive a desired pause number n, and detecting the n most significant pauses in that file and forming speech block boundaries at these n pauses in the speech data.