Speech synthesis using complex spectral modeling

• US 8,280,724 B2
• Filed: 01/31/2005
• Issued: 10/02/2012
• Est. Priority Date: 09/13/2002
• Status: Active Grant

• Current Assignee: Cerence Operating Company (Cerence Inc.)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Chazan, Dan, Hoory, Ron, Shechtman, Slava, Sorin, Alexander, Kons, Zvi
• Primary Examiner(s): Desir, Pierre-Louis
• Assistant Examiner(s): BAKER, MATTHEW H
• Application Number:
  US11/046,911
• Publication Number:
  US 20050131680A1
• Time in Patent Office: 2,801 Days
• Field of Search: 704214-215, 704205-209
• US Class Current: 704/206
• CPC Class Codes: G10L 13/08 Text analysis or generation of parameters for speech synthesis out of text, e.g. grapheme to phoneme translation, prosody generation or stress or intonation determination
• CPC Class Codes: G10L 19/02 using spectral analysis, e.g. transform vocoders or subband vocoders

Abstract

A method for processing a speech signal includes dividing the speech signal into a succession of frames, identifying one or more of the frames as click frames, and extracting phase information from the click frames. The speech signal is encoded using the phase information. Methods are also provided for modeling phase spectra of voiced frames and click frames.

Claims

1. A method for processing a speech signal, comprising using at least one computer programmed to implement:

dividing the speech signal into a succession of frames;

identifying at least one of the frames as an unvoiced click frame;

identifying at least one of the frames as an unvoiced non-click frame;

identifying at least one of the frames as a voiced frame;

calculating one or more parameters of a model of a phase spectrum of the at least one unvoiced click frame;

storing the parameters of the model of the phase spectrum of the at least one unvoiced click frame in a data set;

applying a first method to the at least one unvoiced click frame and to the at least one unvoiced non-click frame to obtain harmonic representations of the at least one unvoiced click frame and the at least one unvoiced non-click frame; and

applying a second method, different from the first method, to the at least one voiced frame to obtain an harmonic representation of the at least one voiced frame,

wherein identifying the at least one of the frames as the at least one unvoiced click frame comprises;

identifying the at least one of the frames as an unvoiced frame; and

processing the at least one unvoiced frame by;

analyzing a probability distribution of the at least one unvoiced frame,

finding a deviation of the probability distribution of the at least one unvoiced frame from a Gaussian distribution, and

identifying the at least one unvoiced frame as the at least one unvoiced click frame if the deviation exceeds a predefined threshold.

2. The method of claim 1, further comprising:

calculating parameters of a model of a phase spectrum of the at least one voiced frame; and

storing the parameters of the model of the phase spectrum of the at least one voiced frame in a data set.

3. The method of claim 2, further comprising:

calculating parameters of models of amplitude spectra of the at least one unvoiced click frame, the at least one unvoiced non-click frame, and the at least one voiced frame, respectively;

storing the parameters of the models of the amplitude spectra of the at least one unvoiced click frame, the at least one unvoiced non-click frame, and the at least one voiced frame in a data set.

4. The method of claim 3, wherein the models of the phase spectra of the at least one unvoiced click frame and the at least one voiced frame are continuous complex phase spectrum models.

5. The method of claim 4, wherein identifying the at least one of the frames as the at least one unvoiced non-click frame comprises determining that the at least one unvoiced non-click frame has a random phase spectrum.

6. The method of claim 2, wherein

calculating the parameters of the model of the phase spectrum of the at least one unvoiced click frame comprises using smooth phase spectrum modeling;

calculating the parameters of the model of the phase spectrum of the at least one voiced frame comprises using smooth phase spectrum modeling; and

using smooth phase spectrum modeling comprises;

using a linear combination of basis functions to model a phase spectrum of a frame, and

aligning and unwrapping respective phases of frequency components of the phase spectrum of the frame before calculating the parameters of the model of the phase spectrum of the frame.

7. The method of claim 2, wherein the model of the phase spectrum of the at least one voiced frame is a time-domain phase spectrum model.

8. The method of claim 1, wherein the model of the phase spectrum of the at least one unvoiced click frame is a continuous complex phase spectrum model.

9. The method of claim 1, wherein processing the at least one unvoiced frame to identify the at least one unvoiced click frame occurs only if a signal level of the at least one unvoiced frame exceeds a predetermined minimum.

10. The method of claim 1, wherein analyzing the probability distribution of the at least one unvoiced frame comprises representing the probability distribution as a histogram of sampled amplitude values of a waveform associated with the at least one unvoiced frame.

11. The method of claim 1, wherein finding the deviation of the probability distribution of the at least one unvoiced frame from a Gaussian distribution comprises estimating an excess of the probability distribution, the excess being equal to a fourth-order centered moment of the probability distribution divided by a square of a second-order centered moment of the probability distribution.

12. The method of claim 1, wherein finding the deviation of the probability distribution of the at least one unvoiced frame from a Gaussian distribution comprises calculating an entropy of the probability distribution.

13. The method of claim 12, wherein the deviation exceeds the predefined threshold if the entropy is less than 2.9.

14. The method of claim 1, wherein

analyzing the probability distribution of an unvoiced frame comprises analyzing a probability distribution of a latter part of the unvoiced frame, and

processing the unvoiced frame further comprises identifying a next frame as an unvoiced click frame if the deviation exceeds the predefined threshold.

15. The method of claim 1, wherein the model of the phase spectrum of the at least one unvoiced click frame represents respective phases of the speech signal at a plurality of frequencies.