Extracting content from speech prosody
AssigneeYeda Research and Development Co. Ltd.
InventorsElisha Moses, Tirza Biron, Dominik Freche, Daniel Baum, Nadav Matalon, Netanel Ehrmann, Eyal Weinreb
Life of this patent
Estimating the term…
Abstract
A prosodic speech recognition engine configured to identify prosodic features and patterns in a speech continuum for the extraction of linguistic content including para-syntactic content, discourse function, information structure, meaning, and speaker sentiment.
Claims20
A method for extracting prosodic content from a speech continuum, performed on a computing device having a processor, memory, one or more code sets comprising instructions for performing the method, the code sets being stored in the memory and executed in the processor, a microphone, and output device, the method comprising:
- capturing, via the microphone, a snippet of a speech continuum of a speaker, the snippet comprising a plurality of words, each word comprising one or more phonemes;
- identifying the words within the snippet;
- for each of the words:
- measuring the duration of each phoneme within the word;
- calculating a word-specific speech rate being an average phoneme rate of phonemes within the word;
- receiving a reference speech rate based on the number of phonemes in, and duration of, at least one of the other words in the snippet;
- and
- calculating a speech rate deviation between the word-specific speech rate and the reference speech rate;
- identifying at least two words characterized by a threshold speech rate deviation;
- denoting, through the output device, each of the at least two words characterized by the threshold speech rate deviation as a boundary of an intonation segment;
- identifying and determining the frequency of a minimum pitch of intra-segment word-strings;
- identifying and determining the frequency of a maximum pitch of the intra-segment word-strings;
- calculating a pitch ratio of the maximum pitch to the minimum pitch;
- comparing the pitch ratio to a gender-specific threshold ratio associated with an inquiry;
- determining an intensity/pitch-product (IP-product) function of the snippet;
- classifying one or more peak IP-product values using a logistic regression model, wherein the logistic regression model is trained to identify a peak IP-product value as being indicative of heightened informativity based on a correlation between the peak IP-product value and one or more prosodic parameters other than the IP-product;
- and
- denoting, through the output device, a word-string comprising one or more words characterized by a pitch ratio exceeding the threshold ratio, and comprising a peak IP-product value classified as being indicative of heightened informativity, as an inquiry.
A prosodic speech recognition engine comprising a microphone, memory, and one or more code sets stored in the memory, an output device, and a processor configured to execute the code sets to:
- capture, via the microphone, a snippet of a speech continuum through the microphone, the snippet comprising a plurality of words, each word comprising one or more phonemes;
- identify the words within the snippet;
- for each of the words:
- measuring the duration of each phoneme within the word;
- calculate a word-specific speech rate being an average phoneme rate of phonemes within the word;
- receive a reference speech rate based on the number of phonemes in, and duration of, at least one of the other words in the snippet;
- and
- calculate a speech rate deviation between the word-specific speech rate and the reference speech rate;
- identify at least two words characterized by a threshold speech rate deviation;
- denote, through the output device, each of the at least two words characterized by the threshold speech rate deviation as a boundary of at least two intonation segments;
- identify and determine the frequency of a minimum pitch of intra-segment word-strings;
- identify and determine the frequency of a maximum pitch of the intra-segment word-strings;
- calculate a pitch ratio of the maximum pitch to the minimum pitch;
- compare the pitch ratio to a gender-specific threshold ratio associated with an inquiry;
- determine an intensity/pitch-product (IP-product) function of the snippet;
- classify one or more peak IP-product values using a logistic regression model, wherein the logistic regression model is trained to identify a peak IP-product value as being indicative of heightened informativity based on a correlation between the peak IP-product value and one or more prosodic parameters other than the IP-product;
- and
- denote, through the output device, a word-string comprising one or more words characterized by a pitch ratio exceeding the threshold ratio, and comprising a peak IP-product value classified as being indicative of heightened informativity, as an inquiry.
Dependencies are inferred from the claim text.