Voice modification detection using physical models of speech production
A computer may train a single-class machine learning using normal speech recordings. The machine learning model or any other model may estimate the normal range of parameters of a physical speech production model based on the normal speech recordings. For example, the computer may use a source-filter model of speech production, where voiced speech is represented by a pulse train and unvoiced speech by a random noise and a combination of the pulse train and the random noise is passed through an auto-regressive filter that emulates the human vocal tract. The computer leverages the fact that intentional modification of human voice introduces errors to source-filter model or any other physical model of speech production. The computer may identify anomalies in the physical model to generate a voice modification score for an audio signal. The voice modification score may indicate a degree of abnormality of human voice in the audio signal.
1 . A computer-implemented method comprising:
extracting, by a computer, a first set of speech parameters representing attributes of one or more normal speech signals, by applying a physical speech model on one or more training audio signals including the one or more normal speech signals;
training, by the computer, a machine-learning model to generate a modified-voice score, by applying the machine-learning model on the first set of speech parameters extracted for the one or more normal speech signals of the one or more training audio signals;
extracting, by the computer, a second set of the speech parameters representing attributes of an inbound speech signal, by applying the physical speech model on an inbound audio signal including the inbound speech signal; and
generating, by the computer, the modified-voice score for the inbound signal by applying the machine-learning model on the second set of the speech parameters and the first set of the speech parameters.
2 . The method according to claim 1 , further comprising determining, by the computer, whether an inbound phone call associated with the inbound audio signal is fraudulent based upon the voice modification score.
3 . The method according to claim 1 , wherein generating the modified-voice score includes comparing, by the computer, the first set of the speech parameters against the second set of the speech parameters to determine one or more similarity scores.
4 . The method according to claim 1 , wherein the machine learning model includes a single-class model trained for detecting one or more audio signal anomalies for the modified-voice score.
5 . The method according to claim 1 , wherein a speech parameter of the speech parameters includes at least one of: a pitch parameter, a formant parameter, a residual parameter, or linear predictive coding (LPC) model order parameter.
6 . The method according to claim 5 , further comprising:
determining, by the computer, a first anomaly score based upon comparing the pitch parameter of the first set of the parameters for the attributes of the normal human speech signals against the pitch parameter of the second set of the parameters for the attributes of the inbound speech signal; and
determining, by the computer, a second anomaly score based upon comparing the formant parameter of the first set of the parameters for the attributes of the normal human speech signals against the formant parameter of the second set of the parameters for the attributes of the inbound speech signal,
wherein the computer determines the modified-voice score based upon each anomaly score.
7 . The method according to claim 1 , wherein the physical speech model applies a source-filter model including a pulse train and white random noise for extracting the speech parameters.
8 . The method according to claim 7 , wherein the source-filter model applies an LPC filter for extracting one or more formant parameters of the speech parameters.
9 . The method according to claim 7 , wherein the source-filter model applies an inverse LPC filter for extracting one or more residual parameters of the speech parameters.
10 . The method according to claim 9 , further comprising generating, by the computer, an LPC filter based upon the one or more residual parameters generated using the inverse LPC filter.
11 . A system comprising:
a computer comprising a processor configured to:
extract a first set of speech parameters representing attributes of one or more normal speech signals, by applying a physical speech model on one or more training audio signals including the one or more normal speech signals;
train a machine-learning model to generate a modified-voice score, by applying the machine-learning model on the first set of speech parameters extracted for the one or more normal speech signals of the one or more training audio signals;
extract a second set of the speech parameters representing attributes of an inbound speech signal, by applying the physical speech model on an inbound audio signal including the inbound speech signal; and
generate the modified-voice score for the inbound signal by applying the machine-learning model on the second set of the speech parameters and the first set of the speech parameters.
12 . The system according to claim 11 , wherein the computer is further configured to determine whether an inbound phone call associated with the inbound audio signal is fraudulent based upon the voice modification score.
13 . The system according to claim 11 , wherein, when generating the modified-voice score, the computer is further configured to compare the first set of the speech parameters against the second set of the speech parameters to determine one or more anomaly scores.
14 . The system according to claim 11 , wherein the machine learning model includes a single-class model trained for detecting one or more audio signal anomalies for the modified-voice score.
15 . The system according to claim 11 , wherein a speech parameter of the speech parameters includes at least one of: a pitch parameter, a formant parameter, a residual parameter, or linear predictive coding (LPC) model order parameter.
16 . The system according to claim 15 , wherein the computer is further configured to:
determine a first anomaly score based upon comparing the pitch parameter of the first set of the parameters for the attributes of the normal human speech signals against the pitch parameter of the second set of the parameters for the attributes of the inbound speech signal; and
determine a second anomaly score based upon comparing the formant parameter of the first set of the parameters for the attributes of the normal human speech signals against the formant parameter of the second set of the parameters for the attributes of the inbound speech signal,
wherein the computer determines the modified-voice score based upon each anomaly score.
17 . The system according to claim 11 , wherein the physical speech model applies a source-filter model including a pulse train and white random noise for extracting the speech parameters.
18 . The system according to claim 17 , wherein the source-filter model applies an LPC filter for extracting one or more formant parameters of the speech parameters.
19 . The system according to claim 17 , wherein the source-filter model applies an inverse LPC filter for extracting one or more residual parameters of the speech parameters.
20 . The system according to claim 19 , wherein the computer is further configured to generate an LPC filter based upon the one or more residual parameters generated using the inverse LPC filter.