Granted Patent B2

US 7,412,377 · App. 10/740,661 · Granted Aug 12, 2008

Abstract

Claims (56)

1. A method of using a computer processor configured for executing operations on encoded computer program instructions stored in computer memory and arranged for generating a spectral feature vector data output, said method comprising the steps of:

accepting at least two input vectors, each input vector including speech or audio or voice spectral features;

creating vectors of ranks via ranking values of the spectral features of each input vector;

creating ordered vectors via arranging the values of each input vector according to rank;

creating a vector of ordered average values via determining the average of corresponding values of the ordered vectors;

creating a vector of ordered average ranks via determining the sum or average of the vectors of ranks;

creating a vector of ordered ranks via ranking the values of the ordered average ranks;

creating a spectral feature vector via employing the rank order represented by the vector of ordered ranks to reorder the vector of ordered average values.

2. The method according to claim 1 , further comprising:

accepting speech or audio or voice input;

said step of accepting at least two input vectors comprises developing input vectors associated with the speech or audio or voice input spectral features;

said method further comprising the step of providing probabilities that each input vector belongs to one or more classes;

said steps of creating a vector of ordered average values and creating a vector of ordered average ranks comprising assigning the probabilities as weights to the input vectors; and

said method further comprising the step of developing a voice model based on the spectral feature vector.

3. The method according to claim 2 , wherein said step of providing probabilities comprises providing probabilities that each vector belongs to particular context-dependent sub-phone units.

4. The method according to claim 2 , further comprising:

generating mel frequency log spectra associated with the speech or audio or voice input;

said step of generating mel frequency log spectra comprising:

segmenting the speech or audio or voice input into a plurality of segments;

applying a Fourier transform and generating a Fourier spectrum for each segment;

binning the Fourier spectrum into channels based on mel frequency; and

determining the logarithm of each channel.

5. The method according to claim 4 , wherein:

said steps of creating vectors of ranks and creating ordered vectors comprise sorting the mel bins to provide a set of ranks and values;

said step of assigning probabilities as weights comprising weighting the sorted values and ranks with the weights and accumulating the weighted ranks and weights as two sets of n-dimensional sums, and developing a total weight for each context-dependent sub-phone unit; and

said steps of creating a vector of ordered average values and creating a vector of ordered average ranks comprising dividing the sums corresponding to each context-dependent sub-phone unit by corresponding total weights to yield a n-dimensional vector of average values and a vector of average ranks.

6. The method according to claim 4 , wherein the input is speech input.

7. The method according to claim 6 , wherein said step of providing probabilities comprises:

determining the probabilities via providing a transcription of the audio input;

expanding words associated with the transcription into phonetic sequences;

joining the phonetic sequences based on a word sequence of the transcription; and

determining context-dependent sub-phonetic units for each phone and employing a speech recognition model to align the result with the segmented audio input.

8. The method accordfng to claim 2 , wherein said at least two input vectors correspond to at least two voice models, whereby said method is adapted for creating from at least two voice models an additional voice model having predetermined characteristics.

9. The method according to claim 2 , wherein the voice model is used as a vector quantization codebook for speech coding.

10. The method according to claim 2 , wherein the voice model is employed in voice transformation via regenerating speech from a sequence of model vectors corresponding to the time sequence of classes corresponding to the audio input.

11. The method according to claim 2 , wherein the voice model is used for voice morphing via:

determining at least one time sequence of classes corresponding to the input speech;

determining a time varying weighted average of spectral feature vectors of

at least two voice models; or

at least one input vector and at least one voice model; and

regenerating speech from the resulting sequence.

12. The method according to claim 2 , wherein the voice model is used for speech synthesis via:

expanding a sequence of words into phonetic sequences;

joining the phonetic sequences based on a word sequence;

expanding the phonetic sequence into one or more classes;

sequencing vectors associated with the voice model according to a class sequence, and

generating speech from the resulting sequence.

13. The method according to claim 1 , wherein said at least two input vectors correspond to at least two spectral models, whereby said method is adapted for creating from at least two spectral models an additional spectral model having predetermined characteristics.

14. The method according to claim 1 , wherein said method is adapted for voice morphing via:

determining a first time sequence of classes corresponding to one instance of input speech;

determining a second time sequence of classes corresponding to another instance of input speech;

determining a time-varying weighted average of the vectors of corresponding classes; and

regenerating speech from the resulting sequence.

15. The method according to claim 1 , wherein:

the set of input vectors is a contiguous set of vectors taken from a time sequence of spectral feature vectors; and

said method further comprises the step of assigning weights for averaging based on the relative position in time of the input spectral feature vectors.

Assignments (10)