System and method for synthesis of speech from provided text

• US 9,911,407 B2
• Filed: 01/14/2015
• Issued: 03/06/2018
• Est. Priority Date: 01/14/2014
• Status: Active Grant

• Current Assignee: Genesys Cloud Services Incorporated
• Original Assignee: Interactive Intelligence Group Incorporated (Genesys Cloud Services Incorporated)
• Inventors: Tan, Yingyi, Ganapathiraju, Aravind, Wyss, Felix Immanuel
• Primary Examiner(s): Pullias, Jesse
• Application Number:
  US14/596,628
• Publication Number:
  US 20150199956A1
• Time in Patent Office: 1,147 Days
• Field of Search: 704257-275
• 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

Abstract

A system and method are presented for the synthesis of speech from provided text. Particularly, the generation of parameters within the system is performed as a continuous approximation in order to mimic the natural flow of speech as opposed to a step-wise approximation of the feature stream. Provided text may be partitioned and parameters generated using a speech model. The generated parameters from the speech model may then be used in a post-processing step to obtain a new set of parameters for application in speech synthesis.

Claims

1. A method for generating parameters in a speech synthesis system, wherein the system comprises a parameter generation module operatively coupled to a speech synthesis module, using a continuous feature stream, for provided text for use in speech synthesis, comprising the steps of:

a) partitioning, by the parameter generation module, said provided text into a sequence of phrases;

b) generating, by the parameter generation module, parameters in a continuous approximation for said sequence of phrases using a speech model; and

c) processing, by the parameter generation module, the generated parameters to obtain an other set of parameters, wherein said other set of parameters comprise at least one clamped delta value and wherein said other set of parameters are utilized in speech synthesis for provided text by the speech synthesis module.

2. The method of claim 1, wherein said partitioning is performed based on linguistic knowledge.

3. The method of claim 1, wherein said speech model comprises a predictive statistical parametric model.

4. The method of claim 1, wherein the generated parameters for the phrases comprise spectral parameters.

5. The method of claim 4, wherein the spectral parameters comprise one or more of the following:

phrase-based spectral parameter values, rate of change of spectral parameters, spectral envelope values, and rate of change of spectral envelope.

6. The method of claim 1, wherein the phrases comprise a grouping of words capable of being separated by at least one of:

linguistic pauses and acoustic pauses.

7. The method of claim 1, wherein the partitioning of said provided text into a sequence of phrases further comprises the steps of:

a) generating a vector based on predicted parameters, wherein said predicted parameters are determined as parameters that represent the text;

b) determining a frame increment value; and

c) determining state of a phrase, wherein

i) if the phrase has started, determining if voicing has started and

1) if voicing has started, adjusting the vector based on parameters of voiced phonemes and restarting step (c);

otherwise,

2) if voicing has ended, adjusting the vector based on parameters of unvoiced phonemes and restarting from step (c);

ii) if the phrase has ended, smoothing the vector and performing a global variance adjustment.

8. The method of claim 1, wherein the generation of the parameters comprises generating a parameter trajectory, which further comprises the steps of:

a) initializing a first element of a generated parameter vector;

b) determining a frame increment value;

c) determining if a linguistic segment is present, wherein

i) if the linguistic segment is not present, determining if voicing has started and

1) if voicing has not started, adjusting the parameter vector based on parameters of voiced phonemes and restarting the process from step (a);

2) if voicing has started, determining if the voicing is in a first frame, wherein, if the voice is in the first frame, a coefficient mean is equal to fundamental frequency, and if the voice is not in the first frame, performing a clamp of the coefficient; and

ii) if the linguistic segment is present, removing abrupt changes of the parameter trajectory, and performing a global variance adjustment.

9. The method of claim 8, wherein step c) i) further comprises the step of determining if voicing has ended, wherein if voicing has not ended, repeating claim 8 from step (a), and if voicing has ended, adjusting the coefficient mean to a desired value and performing long window smoothing on the segment.

10. The method of claim 8, wherein said initializing is performed at time zero.

11. The method of claim 8, wherein said frame increment value comprises a desired integer.

12. The method of claim 11, wherein said desired integer is 1.

13. The method of claim 8, wherein the determining if a frame is voiced comprises examining predicted values for the spectral parameters, wherein a voiced segment comprises valid values.

14. The method of claim 8, wherein the determining if a linguistic segment is present comprises examining a sequence of states for segment partition.

15. The method of claim 1, wherein the generation of parameters comprises generating mel-cepstral parameters, comprising the steps of:

a) initializing a first element of a generated parameter vector;

b) determining a frame increment value;

c) determining if the frame is voiced, wherein;

i) if the segment is unvoiced, applying the mathematical equation;

mcep(i)=(mcep(i−

1)+mcep mean(i))/2;

ii) if the segment is voiced and is a first frame, then applying the mathematical equation;

mcep(i)=(mcep(i−

1)+mcep mean(i))/2; and

iii) if the segment is voiced and is not a first frame, then applying the mathematical equation;

mcep(i)=(mcep(i−

1)+mcep delta(i)+mcep mean(i))/2;

d) determining if a linguistic segment has ended, wherein;

i) if the linguistic segment has ended, removing abrupt changes of the parameter trajectory, and adjusting global variance; and

ii) if the linguistic segment has not ended, repeating the process beginning with step (a).

16. The method of claim 15, wherein said initializing is performed at time zero.

17. The method of claim 15, wherein said frame increment value comprises a desired integer.

18. The method of claim 17, wherein said desired integer is 1.

19. The method of claim 15, wherein the determining if a frame is voiced comprises examining predicted values for the spectral parameters, wherein a voiced segment comprises valid values.