Granted Patent B2

US 12,537,018 · App. 18/217,880 · Granted Jan 27, 2026

Inventors: Daniel Sand (Tel Aviv-Jaffa, IL); Samuel Jefroykin (Tel Aviv-Jaffa, IL); Ilan Kahan (Tel Aviv-Jaffa, IL); Tal Simon (Tel Aviv-Jaffa, IL); Alon Rabinovich (Tel Aviv-Jaffa, IL); Shiri Sadeh-Sharvit (Tel Aviv-Jaffa, IL)

Assignee: ELEOS MENTAL SYSTEMS LTD.

G10L25/63G10L25/30

Abstract

Claims (59)

1 . A method of analyzing speech by at least one processor, the method comprising:

receiving an audio data element representing a discussion;

extracting a first set of audio segments pertaining to speech of a first speaker in the discussion;

analyzing the first set of audio segments by:

applying a band-pass filter on an audio segment of the first set of audio segments, to obtain a filtered version of the audio segment;

applying a Hilbert transform on the filtered version of the audio segment, to obtain a Hilbert envelope;

determining a burst threshold value;

defining at least one audio burst based on the Hilbert envelope and the burst threshold value; and

analyzing the at least one audio burst, to calculate one or more audio features comprising an audio burst Area Under Curve (AUC) defined as an integral between a positive edge of the Hilbert envelope and the burst threshold value; and

applying a machine-learning (ML) model on said one or more audio features, to predict a mental health score, representing a mental condition of the first speaker.

2 . The method of claim 1 , wherein the mental health score represents an expected response of the first speaker in a mental health questionnaire.

3 . The method of claim 1 , further comprising analyzing the first set of audio segments to produce one or more textual n-grams;

analyzing the one or more textual n-grams to produce at least one textual feature, representing a mental condition of the first speaker; and

further applying the ML model on the at least one textual feature, to predict the mental health score.

4 . The method of claim 1 , wherein

said one or more audio features further comprise features selected from a list consisting of: audio burst amplitudes, audio burst duration, and audio burst coefficient of variation (CV).

5 . The method of claim 1 , further comprising:

monitoring the predicted mental health score over time; and

identifying one or more timestamps corresponding to points in the discussion, where the first speaker is suspected to have been in a predefined mental condition, based on said monitoring.

6 . The method of claim 5 , further comprising:

producing a reference data element, comprising a plurality of identified timestamps, corresponding to one or more discussions, and respective predictions of mental health score; and

providing a notification of previous at least one case in which the first speaker is suspected to have been in the predefined mental condition, based on the reference data element.

7 . The method of claim 6 , further comprising:

analyzing the reference data element, in relation to the plurality of identified timestamps; and

producing a recommendation data element, representing a recommendation of treatment, based on said analysis.

8 . The method of claim 6 , further comprising:

extracting a second set of audio segments pertaining to speech of a second speaker in the discussion;

analyzing the second set of audio segments to produce one or more second textual n-grams; and

associating the one or more second textual n-grams to at least one identified timestamp in the reference data element.

9 . A system for analyzing speech, the system comprising: a non-transitory memory device, wherein modules of instruction code are stored, and at least one processor associated with the memory device, and configured to execute the modules of instruction code, whereupon execution of said modules of instruction code, the at least one processor is configured to:

receive an audio data element representing a discussion;

extract a first set of audio segments pertaining to speech of a first speaker in the discussion;

analyze the first set of audio segments by:

applying a band-pass filter on an audio segment of the first set of audio segments, to obtain a filtered version of the audio segment;

applying a Hilbert transform on the filtered version of the audio segment, to obtain a Hilbert envelope;

determining a burst threshold value;

defining at least one audio burst based on the Hilbert envelope and the burst threshold value; and

analyzing the at least one audio burst, to calculate one or more audio features comprising an audio burst Area Under Curve (AUC) defined as an integral between a positive edge of the Hilbert envelope and the burst threshold value; and

apply a machine-learning (ML) model on said one or more audio features, to predict a mental health score, representing a mental condition of the first speaker.

10 . The system of claim 9 , wherein the mental health score represents an expected response of the first speaker in a mental health questionnaire.

11 . The system of claim 9 , wherein the at least one processor is further configured to:

analyze the first set of audio segments to produce one or more textual n-grams;

analyze the one or more textual n-grams to produce at least one textual feature, representing a mental condition of the first speaker; and

apply the ML model on the at least one textual feature, to predict the mental health score.

12 . The system of claim 9 , wherein

said one or more audio features further comprise features selected from a list consisting of: audio burst amplitudes, audio burst duration, and audio burst coefficient of variation (CV).

13 . The system of claim 9 , wherein the at least one processor is further configured to:

monitor the predicted mental health score over time; and

identify one or more timestamps corresponding to points in the discussion, where the first speaker is suspected to have been in a predefined mental condition, based on said monitoring.

14 . The system of claim 13 , wherein the at least one processor is further configured to:

produce a reference data element, comprising a plurality of identified timestamps, corresponding to one or more discussions, and respective predictions of mental health score; and

provide a notification of previous at least one case in which the first speaker is suspected to have been in the predefined mental condition, based on the reference data element.

15 . The system of claim 14 , wherein the at least one processor is further configured to:

perform an analysis of the reference data element, in relation to the plurality of identified timestamps; and

produce a recommendation data element, representing a recommendation of treatment, based on said analysis.

16 . The system of claim 14 , wherein the at least one processor is further configured to:

extract a second set of audio segments pertaining to speech of a second speaker in the discussion;

analyze the second set of audio segments to produce one or more second textual n-grams; and

associate the one or more second textual n-grams to at least one identified timestamp in the reference data element.

Assignments (2)

Continuity (2)

References Cited (10)