Video: Detecting Deception in Speech, Pt. 1: Methods
Learn more about speech technology in academia at the next SpeechTEK conference.
Read the complete transcript of this clip:
Yocheved Levitan: There are lots of challenges involved with detecting deception in text. Firstly, there's a lack of data to study deceptive speech, the way people lie. We don't always have the truth. It's very hard to know what people are comparing things with, and also when there are studies, it's very difficult to compare them with the real world.
Also, there are differences in the way people lie across cultures and across gender, et cetera. So this is a more recent approach presented in the PhD thesis of Sarah Ita Levitan, who happens to be my sister, and she studied deception detection in the area through the vehicle of speech.
Some notable features of this work is that it offers a large-scale corpus for studying deception, multi-domain, across cultures, across gender. Also, financial incentive for people who lie well and for people who deceive well. And methods for automatically extracting features and machine-learning approaches for classifying deceptive speech.
Three main ideas are identifying cues to deception, studying speaker variability in the way people lie, and using this information to classify speech. A little bit more detail about the corpus known as the CXD corpus, or the Columbia Cross-Cultural Deception Corpus, this is an overview of the experiment where they, it's an extension of the false resume paradigm, also known as like a lying game where participants came into the lab, and one interviewed another, asking questions where one participant was told to lie.
Another was supposed to detect whether they were saying the truth or not. And they were offered money for performing well. It also included personality scores and a baseline sample of their speech to compare with. Here's an example of a questionnaire that they were given where, for half the examples, they were told to fabricate a response, and then when their partner asked them, they would say the false answer. For example, have you ever, do you have allergies to any foods? For that example, they would lie, and then they would have to try to get away with it.
So just some other features of this corpus, 340 subjects, over 120 hours of speech, different cultures, different gender, also labeled with global and local deception. They studied different segments of speech, different units of speech, inter-pausal units, which is pause-free speech. Turns is just one speaker but may include pauses. Question responses are those that directly follow a prompt like, do you have allergies, so the direct response to that. And then question chunks are the total response to a given question.
So once they collected all that data, they extracted different feature sets. Acoustic-prosodic are the speech features. Then, linguistic deception indicators are different features that are explicitly indicative of deception like laughter, hesitation, et cetera. LIWC features are different dimensions of speech like formality and punctuation. And complexity studies the syntax of the speech, how complex it is. In total, there were 152 features studied, three different groups. For example, for acoustic-prosodic, these are the pitch features, different measure for each. Intensity is the volume.
Also there are different measures, and voice quality. We know that people reveal themselves to the world through their words, so it's not only useful to look at the speech sample, but to also look at the content. And that's why we use LIWC to identify different features in the text, which were obtained through transcribing the audio. And these different dimensions can give us some insight into the psychological cues of the speech.
Related Articles
Greg Stack, Vice President, Speech-Soft Solutions, will be presenting at SpeechTEK 2020. Conference co-chair, Jim Larson, caught up with him to talk about "Blazing a Trail to Successful AI Migration."
05 Nov 2019
Nuance Communications Senior Manager, Commercial Security Strategy Roanne Levitt explains how Nuance's Conversation Print addresses speech spoofing in this clip from her panel an SpeechTEK 2019.
01 Nov 2019
Nuance Communications Senior Manager, Commercial Security Strategy Roanne Levitt and ID R&D VP of Sales John Amein discuss the essential requirements for complying with privacy regulations in this clip from their panel at SpeechTEK 2019.
25 Oct 2019
Pindrop Director of Product Marketing Ben Cunningham discusses how liveness detection, synthetic speech detection, and other deep voice biometric technologies can fight the threat deepfakes pose in this clip from a panel at SpeechTEK 2019.
18 Oct 2019
The latest advances in speech-to-text technology enable analytics software to deliver the most accurate and complete customer insights. To do so, robust ASR solutions must be able to do five things.
16 Oct 2019
Interactions LLC NLP Scientist Yocheved Levitan outlines the findings of a recent academic study on detecting deceptive speech and how to distinguish it from truthful speech in this clip from her presentation at SpeechTEK 2019.
11 Oct 2019