Granted Patent B2

US 12,475,913 · App. 17/989,978 · Granted Nov 18, 2025

Inventors: Guy G. Michaeli (Seattle, WA); Timothy D. Cline (Gainesville, VA); Stephen J. Green (Burlington, MA); Serge Le Huitouze (Rennes, FR); Matthew T. Gerdes (San Diego, CA); Guang Chao Wang (San Diego, CA); Kenny C. Gross (Escondido, CA)

Assignee: Oracle International Corporation

G10L25/63G10L15/02G10L25/03

Abstract

Claims (93)

1 . A non-transitory computer-readable medium that includes stored thereon computer-executable instructions that when executed by at least a processor of a computer system cause the computer system to:

detect a human voice as an anomaly in audio content of a radio signal that includes the human voice and static noise from the radio signal;

extract data that represents the human voice from the audio content by:

(a) generating second audio content that is estimated for the audio content by an extraction ML model, wherein the extraction ML model is trained on second static noise of radio signals that do not include voice to approximate the static noise overlaying the human voice in the audio content; and

(b) generating a residual surface from (i) the audio content that includes the human voice and the static noise and (ii) the second audio content that is estimated for the audio content by the extraction ML model, wherein the data that represents the human voice is the residual surface;

determine a cumulative mean absolute error from the residual surface;

recognize speech in the data that represents the human voice to transform the data that represents the human voice into vocal metrics;

generate feature scores that represent features of the recognized speech based at least in part on the vocal metrics;

classify the human voice as either a hoax distress call or an authentic distress call based on the feature scores;

include in an alert a detection confidence factor that the human voice detected is actually a human voice, wherein the detection confidence factor is based on the cumulative mean absolute error; and

present the alert that indicates that the human voice is one of the hoax distress call or the authentic distress call.

2 . The non-transitory computer-readable medium of claim 1 , wherein the instructions to generate feature scores that represent features of the recognized speech based at least in part on the vocal metrics further cause the computer system to:

generate a feature score for tone or inflection based at least in part on one or more of the vocal metrics of volume, pitch, or pace for the recognized speech, wherein the feature score for tone or inflection indicates an extent to which the tone or inflection of the human voice is consistent with a genuine distress call, and wherein the feature score for tone or inflection is one of the feature scores.

3 . The non-transitory computer-readable medium of claim 1 , wherein the instructions to generate feature scores that represent features of the recognized speech based at least in part on the vocal metrics further cause the computer system to:

generate the feature score for phrasing or sentiment based on one or more of words or an arrangement of words in a text of the recognized speech, wherein the feature score for phrasing or sentiment indicates an extent to which the phrasing or sentiment in the text of the recognized speech is consistent with a genuine distress call, and wherein the feature score for phrasing or sentiment is one of the feature scores.

4 . The non-transitory computer-readable medium of claim 1 , further comprising instructions that when executed by at least the processor cause the computer system to:

generate a residual surface from the audio content that includes the human voice and second audio content that does not include human voice;

apply a first weight to residuals of the residual surface in response to the classification of the human voice as the hoax distress call, or apply a second weight to the residuals of the residual surface in response to the classification of the human voice as the authentic distress call;

determine a cumulative mean absolute error from the weighted residual surface; and

include in the alert a hoax confidence factor that the human voice is correctly classified as a hoax, wherein the hoax confidence factor is based on the cumulative mean absolute error.

5 . The non-transitory computer-readable medium of claim 1 , wherein the instructions to detect the human voice as an anomaly in the audio content of the radio signal further cause the computer system to:

divide a spectrum of the audio content into a set of frequency bins;

select a subset of the frequency bins that are most likely to include human voice distress calls; and

detect the human voice in the subset of the frequency bins with a first machine learning model that has been trained to identify human voice in the subset of frequency bins as an anomaly.

6 . The non-transitory computer-readable medium of claim 1 , wherein the instructions to extract the data that represents the human voice from the audio content further cause the computer system to:

in response to detecting the human voice, begin recording the audio content;

in response to detecting that the human voice has ceased in the audio content for a pre-determined amount of time, end recording the audio content; and

extract the data that represents the human voice from the audio content with a machine learning model that has been trained to identify human voice in the audio content as an anomaly.

7 . A computer-implemented method, comprising:

receiving a suspected distress call in audio content of a radio signal that includes first static noise from the radio signal;

generating second audio content that is estimated for the audio content by an extraction ML model, wherein the extraction ML model is trained on second static noise of radio signals that do not include voice to approximate the first static noise in the audio content;

generating residuals between (i) the audio content that includes the suspected distress call and (ii) the second audio content that contains second radio frequency noise and does not contain human voice to extract audio of the suspected distress call;

generating a detection confidence factor that the suspected distress call was correctly detected based on the residuals;

converting the audio of the suspected distress call into text of the call and vocal metrics of the call;

generating feature scores representing features of the suspected distress call based on the text and vocal metrics;

generating a classification, by a machine learning model, of the suspected distress call as either a hoax distress call or an authentic distress call based on the feature scores, wherein the machine learning model is trained based on feature scores from a set of known authentic distress calls and a set of known hoax distress calls;

and

generating an alert that indicates the classification of the suspected distress call and the detection confidence factor.

8 . The computer-implemented method of claim 7 , further comprising:

generating residuals between first data that represents first audio of the suspected distress call and second data that represents second audio that contains radio frequency noise and does not contain a distress call;

weighting the residuals with a first weight where the suspected distress call is classified as a hoax or weighting the residuals with a second weight where the suspected distress call is classified as authentic; and

generating a hoax confidence factor that the classification is correct based on the weighted residuals.

9 . The computer-implemented method of claim 7 , wherein the generating feature scores representing features of the suspected distress call based on the text and vocal metrics further comprises:

generating a feature score for tone based at least in part on a vocal metric of pitch for the suspected distress call;

generating a feature score for inflection based at least in part on a vocal metric of volume for the suspected distress call;

generating a feature score for phrasing based at least in part on analysis of a sequence of words in the suspected distress call; and

generating a feature score for sentiment based at least in part on an extent to which an emotion is conveyed in the suspected distress call.

10 . The computer-implemented method of claim 7 , wherein receiving a suspected distress call in audio format further comprises:

selecting a subset of frequency bins that are most likely to include human voice distress calls based on analysis of historical authentic distress calls;

detecting the suspected distress call in the subset of the frequency bins; and

in response to detecting the human voice in the subset of the frequency bins, begin recording the suspected distress call.

11 . The computer-implemented method of claim 7 , wherein in response to the classification of the suspected distress call to be the authentic distress call, the alert causes an initiation of a rescue.

12 . The computer-implemented method of claim 7 , further comprising:

before converting the suspected distress call into text, iteratively training a machine learning model with audio noise to identify human voice as an anomaly in noise until a threshold is satisfied for a word error rate of text converted from audio extracted as the anomaly; and

extracting data that represents the suspected distress call from audio content by detecting the suspected distress call as an anomaly in the audio content using the trained machine learning model.

13 . A computing system, comprising:

at least one processor;

a radio receiver configured to scan a distress frequency; and

a non-transitory computer readable medium including instructions stored thereon that when executed by at least the processor cause the computing system to:

autonomously discriminate a human voice from radiofrequency noise from a radio signal in audio content that is received from the radio receiver;

extract data that represents the human voice from the audio content using an anomaly detection model that has been iteratively trained to reduce a word error rate in speech recognition by:

(a) generating second audio content that is estimated for the audio content by the anomaly detection model, wherein the anomaly detection model is trained on second radiofrequency noise of radio signals that do not include voice to approximate the radiofrequency noise overlaying the human voice in the audio content; and

(b) generating a residual surface from (i) the audio content that includes the human voice and the radiofrequency noise and (ii) the second audio content that is estimated for the audio content by the anomaly detection model, wherein the data that represents the human voice is the residual surface;

determine a cumulative mean absolute error from the residual surface;

recognize speech in the data that represents the human voice to transform the data that represents the human voice into vocal metrics;

generate feature scores that represent features of the recognized speech including at least one of tone or inflection, wherein the feature scores are generated based at least in part on the vocal metrics;

classify the human voice as either a hoax distress call or an authentic distress call based on the feature scores;

include in an alert a detection confidence factor that the human voice extracted is actually a human voice, wherein the detection confidence factor is based on the cumulative mean absolute error; and

present the alert indicating that the human voice is one of the hoax distress call or the authentic distress call with a hoax confidence factor that the human voice is correctly classified as the hoax distress call or the authentic distress call.

14 . The computing system of claim 13 , wherein the instructions further cause the computing system to:

generate a residual surface from the audio content that includes the human voice and second audio content that does not include human voice;

apply a first weight to residuals of the residual surface in response to the classification of the human voice as the hoax distress call, or apply a second weight to the residuals of the residual surface in response to the classification of the human voice as the authentic distress call; and

determine a cumulative mean absolute error from the residual surface;

wherein the hoax confidence factor is based on the cumulative mean absolute error.

15 . The computing system of claim 13 , wherein the instructions further cause the computing system to iteratively train the anomaly detection model with audio noise to identify human voice as an anomaly in noise until a threshold is satisfied for the word error rate of text recognized from the data that represents the human voice extracted as the anomaly.

16 . The computing system of claim 13 , wherein the instructions to extract the human voice from the audio content of the radio signal further cause the computing system to:

divide a spectrum of the audio content into a set of frequency bins;

select a subset of the frequency bins that are most likely to include human voice distress calls;

detect the human voice in the subset of the frequency bins with a first machine learning model that has been trained to identify human voice in the subset of frequency bins as an anomaly;

in response to detecting the human voice in the subset of the frequency bins, begin recording time series signals for the set of frequency bins;

detect that the human voice has ceased in the subset of the frequency bins with a second machine learning model; and

in response to detecting that the human voice has ceased for a pre-determined amount of time, end recording the time series signals for the set of frequency bins.

17 . The computing system of claim 16 , wherein the instructions to extract the human voice from the recorded time series signals further cause the computing system to:

generate a time series signal from each bin in the set of frequency bins to produce a set of time series signals;

generate an estimated time series signal for each time series signal based on the other time series signals in the set of time series signals;

generate residuals between each time series signal in the set of time series signals and the estimated time series signal estimated for the time series signal to produce a set of residual time series signals; and

provide the set of residual time series signals as the data that represents the human voice extracted from the audio content, wherein the data that represents the human voice that is extracted from the audio content is denoised based on the estimated time series signals.

18 . The computer-implemented method of claim 7 , wherein when the human voice is an authentic distress call, the method further comprises automatically dispatching rescue resources by an electronic message to rescuers.

19 . The computer-implemented method of claim 7 , wherein the human voice falls below a noise floor of the radio signal.

20 . The computer-implemented method of claim 7 , wherein the extraction ML model is a multivariate ML anomaly detection model, the method further comprising:

select a set of audio frequencies that are associated with human voice;

configure the extraction ML model to receive the set of audio frequencies from the second static noise of radio signals that do not include voice as multivariate training input to the extraction ML model; and

iteratively (a) train the extraction ML model on the set of audio frequencies from the second static noise to identify human voice as an anomaly in the second static noise, and (b) test the extraction ML model on the set of audio frequencies from training audio content that includes test human voice and test static noise from a test radio signal, until word error rate in speech recognition of the test voice identified as an anomaly satisfies a threshold.

Assignments (1)

Continuity (2)

References Cited (125)