Obfuscated speech synthesis

1. A method for synthesizing a speech signal, comprising the steps of:

obtaining a speech sequence input signal comprising semantic content corresponding to a speaker's utterance;

analyzing the input speech sequence signal to obtain a first sequence of feature vectors for the input speech sequence signal;

synthesizing a second sequence of feature vectors different from the first sequence of feature vectors and based on the first sequence of feature vectors, wherein the second sequence of feature vectors retains all relevant vocal characteristics suitable for speaker recognition of the speaker's voice, wherein the second sequence of feature vectors comprises no meaningful semantic information related to the input speech sequence signal;

generating an excitation signal based on the obtained input speech sequence signal;

obfuscating the semantic content of the synthesized speech signal by filtering the excitation signal based on the second sequence of feature vectors wherein the speaker's vocal characteristics are retained to remain suitable for speaker recognition, wherein the semantic content comprises at least a portion of morphemes uttered, and wherein the vocal characteristics comprise at least a portion of phonemes uttered;

wherein synthesizing the second sequence of feature vectors is based on:

determining autoregressive Gaussian Mixture Model parameters for training speech data provided by the speaker;

determining the most likely sequence of acoustic states for the input speech sequence signal based on the autoregressive Gaussian Mixture Model;

shuffling the most likely sequence of acoustic states to obtain a shuffled sequence of acoustic states; and

determining the second sequence of feature vectors as the most likely sequence of feature vectors corresponding to the shuffled sequence of acoustic states based on the determined autoregressive Gaussian Mixture Model parameters.

2. The method according to claim 1 , wherein the most likely sequence of acoustic states for the input speech sequence signal is determined by the Viterbi algorithm.

3. The method according to claim 1 , wherein the autoregressive Gaussian Mixture Model parameters are determined based on the assumption that the mean vectors are linear functions of past sequences of feature vectors.

4. The method according to claim 1 , wherein the autoregressive Gaussian Mixture Model parameters are determined by the Expectation Maximization approach.

5. The method according to claim 1 , wherein the second sequence of feature vectors is obtained as a linear function of an expectation value of past sequences of feature vectors.

6. The method according to claim 1 , wherein each of the feature vectors of the first and second sequences of feature vectors comprises MEL frequency spectrum coefficients or Linear Predictive Coefficients.

7. The method of claim 1 , further comprising the step of analyzing the synthesized speech signal wherein the semantic content is obfuscated for matching with stored templates obtained for speech utterances of the speaker.