Method for the selection of synthesis units

• US 8,195,463 B2
• Filed: 10/22/2004
• Issued: 06/05/2012
• Est. Priority Date: 10/24/2003
• Status: Expired due to Fees

• Current Assignee: Thales SA
• Original Assignee: Thales SA
• Inventors: Capman, François, Padellini, Marc
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): Serrou, Abdelali
• Application Number:
  US10/970,731
• Publication Number:
  US 20050137871A1
• Time in Patent Office: 2,783 Days
• Field of Search: 704/258, 704/268, 704500-504
• US Class Current: 704/258
• CPC Class Codes: G10L 13/06 Elementary speech units used in speech synthesisers; Concatenation rules
• CPC Class Codes: G10L 19/0018 Speech coding using phonetic or linguistical decoding of the source; Reconstruction using text-to-speech synthesis

Abstract

A method for the selection of synthesis units of a piece of information that can be decomposed into synthesis units, comprises at least the following steps for a considered information segment: determining the mean fundamental frequency value F0 for the information segment considered; selecting a sub-set of synthesis units defined as being the sub-set whose mean pitch values are the closest to the pitch value F0; applying one or more proximity criteria to the selected synthesis units to determine a synthesis unit representing the information segment.

Claims

1. A method for selecting synthesis units of a piece of information, the information being a speech segment to be encoded that can be decomposed into synthesis units for a considered information segment, said method comprising the steps:

determining a mean fundamental frequency value F0 for the information segment considered,

identifying a sub-set of the dictionary corresponding to the frequency F0,

wherein said sub-set of the dictionary has N synthesis units,

selecting a sub-set of P synthesis units in said N synthesis units of said identified sub-set of the dictionary, said sub-set of P synthesis units defined as being the closest units whose mean pitch values are the closest to the pitch value F0, the rest of the processing being done on the pitch profiles associated with said P units, with P inferior to N, P=2^nbits, and

applying one or more proximity criteria to the selected synthesis units to determine a synthesis unit representing the information segment.

2. The method according to claim 1, wherein the criteria used as proximity criteria are the fundamental frequency or pitch, the spectral distortion, and/or the energy profile and a step is executed for the combining of the criteria used in order to determine the representative synthesis unit.

3. The method according to claim 1, wherein, for a speech segment to be encoded, the reference pitch is obtained from a prosody generator.

4. The method according to claim 2, wherein the estimation of the criterion of similarity for the profile of the pitch comprises the following steps:

A1) the selection, in the identified sub-set of the dictionary, of the synthesis units and from the mean value of the pitch, of the N closest units in the sense of the criterion of the mean pitch,

A2) the temporal aligning of the N profiles with that of the segment to be encoded,

A3) the computing of N measurements of similarity between the N aligned pitch profiles and the pitch profile of the speech segment to be encoded to obtain the N coefficients of similarity {rp(1), rp(2), . . . rp(N)}.

5. The method according to claim 4, wherein the temporal alignment is a temporal alignment obtained by DTW (dynamic time warp) programming or an alignment by linear adjustment of the lengths.

6. The method according to claim 4, wherein the measurement of similarity is a standardized intercorrelation measurement.

7. The method according to claim 2, wherein the estimation of similarity for the energy profile comprises the following steps:

A4) the determining of the profiles of evolution of energy for the N selected units according to a criterion of proximity of the mean pitch;

A5) the temporal aligning of the N profiles with that of the segment to be encoded;

A6) the computing of N measurements of similarities, between the N profiles of aligned energy values and the energy profile of the speech segment to be encoded to obtain the N coefficients of similarity {re(1), re(2), . . . , re(N)}.

8. The method according to claim 7, wherein the temporal alignment is a temporal alignment obtained by DTW (dynamic time warp) programming or an alignment by linear adjustment of the lengths.

9. The method according to claim 7, wherein the measurement of similarity is a standardized intercorrelation measurement.

10. The method according to claim 2, wherein the estimation of the criterion of similarity for the spectral envelope comprises the following steps:

A7) the temporal aligning of the N profiles with that of the segment to be encoded,

A8) the determining of the profiles of evolution of the spectral parameters for the N selected units according to a criterion of proximity of the mean pitch,

A9) the computing of N measurements of similarities, between the spectral sequence of the segment to be encoded and the N spectral sequences extracted from the selected synthesis units to obtain the N coefficients of similarity {rs(1), rs(2), . . . , rs(N)}.

11. The method according to claim 10, wherein the temporal alignment is a temporal alignment obtained by DTW (dynamic time warp) programming or an alignment by linear adjustment of the lengths.

12. The method according to claim 10, wherein the measurement of similarity is a standardized intercorrelation measurement.

13. The method according to claim 10, wherein the measurement of similarity is a measurement of spectral distance.

14. The method according to claim 10, wherein the step A9) comprises a step in which the set of spectra of a same segment is averaged and wherein the measurement of similarity is a measurement of intercorrelation.

15. The method according to claim 10, wherein the criterion of spectral distortion is computed on harmonic structures re-sampled at constant pitch or re-sampled at the pitch of the segment to be encoded, after interpolation of the initial harmonic structures.

16. The method according to claim 1, comprising a step of encoding and/or a step of correction of the pitch by modification of the synthesis profile.

17. The method according to claim 16, wherein step of encoding and/or correction of the pitch may be a linear transformation of the profile of the original pitch.

18. The use of the method according to claim 1 used for the selection and/or the encoding of synthesis units for a speech encoder working at very low bit rates.

19. The method according to claim 1, wherein said dictionary is divided into 64 sub-classes, where each sub-class includes the synthesis units that are temporally preceded by a segment belong to a same class.

20. The method according to claim 1, wherein a bit rate associated with the encoding scheme shows that

Index of class on 6 bits (64 classes)

Index of the units selected on 5 bits (32 units per sub-class)

N=32 corresponding to the number of the F_0moyen.