Method and apparatus for differentiating between human and electronic speaker for voice interface security
A system for distinguishing between a human voice generated command and an electronic speaker generated command is provided. An exemplary system comprises a microphone array for receiving an audio signal collection, preprocessing circuitry configured for converting the audio signal collection into processed recorded audio signals, energy balance metric determination circuitry configured for calculating a final energy balance metric based on the processed recorded audio signals, and energy balance metric evaluation circuitry for outputting a command originator signal based at least in part on the final energy balance metric.
1. A system for distinguishing between a human voice generated command and an electronic speaker generated command, the system comprising:
a microphone array for receiving an audio signal collection;
preprocessing circuitry configured for converting the audio signal collection into processed recorded audio signals;
energy balance metric determination circuitry configured for calculating a final energy balance metric based on the processed recorded audio signals,
wherein calculating the final energy balance metric comprises:
applying a sliding window to the recorded audio signals to produce a plurality of recorded audio signal windows;
for each recorded audio signal window:
calculating a fast Fourier transform (FFT) for the recorded audio signal window to produce a recorded audio signal window FFT;
cropping the recorded audio signal window FFT down such that a resulting cropped recording audio signal window FFT comprises frequencies within a defined frequency range;
integrating over the cropped recording audio signal window FFT to produce a spectra energy curve;
normalizing the spectral energy curve such that an area underneath the spectral energy curve is equal to one; and
defining a cutoff value based on the normalized spectral energy curve, wherein the cutoff value represents a separate of sub-bass and bass regions and defines an energy balance metric for the recorded audio signal window;
fitting the energy balance metrics for each of the recorded audio signal windows to a normal distribution to produce a normalized energy balance metric distribution; and
assigning a median value from the normalized energy balance metric distribution to the final energy balance metric; and
energy balance metric evaluation circuitry for outputting a command originator signal based at least in part on the final energy balance metric.
2. The system of claim 1 , wherein the preprocessing circuitry comprises circuitry for microphone equalizing, amplitude normalization, and noise filtering.
3. The system of claim 1 , wherein the spectral energy curve represents total energy of the associated recorded audio signal in the defined frequency range.
4. The system of claim 1 , wherein the defined frequency range is between 20 Hz and 250 Hz.
5. The system of claim 1 , wherein the FFT is calculated with 4096 output points for each recorded audio signal window.
6. The system of claim 1 , wherein a size of a recorded audio signal window is 0.1 seconds.
7. The system of claim 6 , wherein the recorded audio signal windows do not overlap one another.
8. The system of claim 1 , wherein normalizing the spectral energy curve results in a value at any point along the spectral energy curve equal to a cumulative distribution function.
9. The system of claim 1 , wherein the energy balance metric for a recorded audio signal is equivalent to the following equation:
energy
balance
metric
=
E
Sub
-
bass
Region
E
Total
Evaluated
Region
(
1
)
where E Sub-bass Region represents energy accumulated in the sub-bass region and E Total Evaluated Region is energy accumulated in the defined frequency range.
10. The system of claim 1 , wherein fitting the energy balance metrics for each of the recorded audio signal windows to a normal distribution to produce a normalized energy balance metric distribution comprises removing outliers based on a skewdness of the energy balance metrics.
11. The system of claim 1 , wherein the command original signal indicates whether the audio signal collection originated from a human voice or from an electronic speaker.
12. A method for distinguishing between a human voice generated command and an electronic speaker generated command, comprising:
receiving an audio signal collection;
converting the audio signal collection into processed recorded audio signals;
calculating a final energy balance metric based on the processed recorded audio signals;
wherein calculating the final energy balance metric comprises:
applying a sliding window to the recorded audio signals to produce a plurality of recorded audio signal windows;
for each recorded audio signal window:
calculating a fast Fourier transform (FFT) for the recorded audio signal window to produce a recorded audio signal window FFT;
cropping the recorded audio signal window FFT down such that a resulting cropped recording audio signal window FFT comprises frequencies within a defined frequency range;
integrating over the cropped recording audio signal window FFT to produce a spectra energy curve;
normalizing the spectral energy curve such that an area underneath the spectral energy curve is equal to one; and
defining a cutoff value based on the normalized spectral energy curve, wherein the cutoff value represents a separate of sub-bass and bass regions and defines an energy balance metric for the recorded audio signal window;
fitting the energy balance metrics for each of the recorded audio signal windows to a normal distribution to produce a normalized energy balance metric distribution; and
assigning a median value from the normalized energy balance metric distribution to the final energy balance metric; and
outputting a command originator signal based at least in part on the final energy balance metric.
13. The method of claim 12 , wherein converting the audio signal collection into processed recorded audio signals comprises microphone equalizing, amplitude normalization, and noise filtering.
14. The method of claim 12 , wherein the spectral energy curve represents total energy of the associated recorded audio signal in the defined frequency range.
15. The method of claim 12 , wherein the defined frequency range is between 20 Hz and 250 Hz.
16. The method of claim 12 , wherein the FFT is calculated with 4096 output points for each recorded audio signal window.
17. The method of claim 12 , wherein a size of a recorded audio signal window is 0.1 seconds.
18. The method of claim 12 , wherein the energy balance metric for a recorded audio signal is equivalent to the following equation:
energy
balance
metric
=
E
Sub
-
bass
Region
E
Total
Evaluated
Region
(
1
)
where E Sub-bass Region represents energy accumulated in the sub-bass region and E Total Evaluated Region 1 S energy accumulated in the defined frequency range.
19. The method of claim 12 , wherein fitting the energy balance metrics for each of the recorded audio signal windows to a normal distribution to produce a normalized energy balance metric distribution comprises removing outliers based on a skewdness of the energy balance metrics.
20. The method of claim 12 , wherein the command original signal indicates whether the audio signal collection originated from a human voice or from an electronic speaker.