A growing body of work on adversarial examples has identified that for machine-learning (ML) systems that operate on high-dimensional data, for nearly every natural input there exists a small perturbation of the point that will be misclassified by the system, posing a threat to its deployment in certain critical settings(Behzadan and Munir, 2017; Chen et al., 2018; Evtimov et al., 2018; Huang et al., 2017; Kurakin et al., 2016; Papernot et al., 2017, 2016). More broadly, the susceptibility of ML systems to adversarial examples has prompted a re-examination of whether current ML systems are truly learning or if they are assemblages of tricks that are effective yet brittle and easily fooled (Samhita and Gross, 2013)
. Implicit in this line of reasoning is the assumption that instances of ”real" learning, such as human cognition, yield extremely robust systems. Indeed, at least in computer vision, human perception is regarded as the gold-standard for robustness to adversarial examples.
Evidently, humans can be fooled by a variety of illusions, whether they be optical, auditory, or other; and there is a long line of research from the cognitive science and psychology communities investigating these (Hillis et al., 2002). In general, however, these illusions are viewed as isolated examples that do not arise frequently, and which are far from the instances encountered in everyday life.
In this work, we attempt to understand how susceptible humans’ perceptual systems for natural speech are to carefully designed “adversarial attacks.” We investigate the density of certain classes of illusion, that is, the fraction of natural language utterances whose comprehension can be affected by the illusion. Our study centers around the McGurk effect, which is the well-studied phenomenon by which the perception of what we hear can be influenced by what we see (McGurk and MacDonald, 1976). A prototypical example is that the audio of the phoneme “baa,” accompanied by a video of someone mouthing “vaa”, can be perceived as “vaa” or “gaa” (Figure 1). This effect persists even when the subject is aware of the setup, though the strength of the effect varies significantly across people and languages and with factors such as age, gender, and disorders (Bastien-Toniazzo et al., 2010; Delbeuck et al., 2007; Irwin et al., 2006; Mongillo et al., 2008; Norrix et al., 2006, 2007; Sekiyama, 1997; Sekiyama et al., 2014; Youse et al., 2004).
A significant density of illusionable instances for humans might present similar types of security risks as adversarial examples do for ML systems. Auditory signals such as public service announcements, instructions sent to first responders, etc., could be targeted by a malicious agent. Given only access to a screen within eyesight of the intended victims, the agent might be able to significantly obfuscate or alter the message perceived by those who see the screen (even peripherally).
1.1 Related work
Illusionable instances for humans are similar to adversarial examples for ML systems. Strictly speaking, however, our investigation of the density of natural language for which McGurk illusions can be created, is not the human analog of adversarial examples. The adversarial examples for ML systems are datapoints that are misclassified, despite being extremely similar to a typical datapoint (that is correctly classified). Our illusions of misdubbed audio are not extremely close to any typically encountered input, since our McGurk samples have auditory signals corresponding to one phoneme/word and visual signals corresponding to another. Also, there is a compelling argument for why the McGurk confusion occurs, namely that human speech perception is bimodal (audio-visual) in nature when lip reading is available(Besle et al., ; Summerfield, 1992).
To the best of our knowledge, prior to our work, there has been little systematic investigation of the extent to which the McGurk effect, or other types of illusions, can be made dense in the set of instances encountered in everyday life. The closest work is Elsayed et al. (2018)
, where the authors demonstrate that some adversarial examples for computer vision systems also fool humans when humans were given less than a tenth of second to view the image. However, some of these examples seem less satisfying as the perturbation acts as a pixel-space interpolation between the original image and the “incorrect” class. This results in images that are visually borderline between two classes, and as such, do not provide a sense of illusion to the viewer. In general, researchers have not probed the robustness of human perception with the same tools, intent, or perspective, with which the security community is currently interrogating the robustness of ML systems.
2 Problem setup
For the McGurk effect, we attempt an illusion for a language token (e.g. phoneme, word, sentence) by creating a video where an audio stream of is visually dubbed over by a person saying . We stress that the audio portion of the illusion is not modified and corresponds to a person saying . The illusion affects a listener if they perceive what is being said to be if they watched the illusory video whereas they perceive if they had either listened to the audio stream without watching the video or had watched the original unaltered video, depending on specification. We call a token illusionable if an illusion can be made for the token that affects the perception of a significant fraction of people.
In Section 3, we analyze the extent to which the McGurk effect can be used to create illusions for phonemes, words, and sentences, and analyze the fraction of natural language that is susceptible to such illusionability. We thereby obtain a lower bound on the density of illusionable natural speech.
We find that 1) a significant fraction of words that occur in everyday speech can be turned into McGurk-style illusions, 2) such illusions persist when embedded within the context of natural sentences, and in fact affect a significant fraction of natural sentences, and 3) the illusionability of words and sentences can be predicted using features from natural language modeling.
3 McGurk experiments
3.1 Phoneme-level experiments
We began by determining which phoneme sounds can be paired with video dubs of other phonemes to effect a perceived phoneme that is different from the actual sound. We created McGurk videos for all vowel pairs preceded with the consonant /n/ as well as for all consonant pairs followed by the vowel /a/ spoken by a speaker. There are 20 vowel phonemes and 24 consonant phonemes in American English although // and // are redundant for our purposes.111Refer to www.macmillanenglish.com/pronunciation/phonemic-chart-in-american-english/ for a list of the phoneme symbols of the International Phonetic Alphabet that are used in American English. Based on labels provided by 10 individuals we found that although vowels were not easily confused, there are a number of illusionable consonants. We note that the illusionable phoneme pairs depend both on the speaker and listener identities.
Given Table 1 of illusionable phonemes, the goal was then to understand whether these could be leveraged within words or sentences; and if so, the fraction of natural speech that is susceptible.
|Phoneme||Lip Movement||Perceived Sound|
|b/||/w/||/v/, /f/, /p/|
|f/||/z/||/T/, /t/, /b/|
|m/||/D/||/nT/, /n/, /ml/|
|T/||/v/||/d/, /k/, /t/, /f/|
3.2 Word-level experiments
We sampled 200 unique words (listed in Table 2) from the 10,000 most common words in the Project Gutenberg novels in proportion to their frequency in the corpus. The 10k words collectively have a prevalence of 80.6% in the corpus. Of the 200 sampled words, 147 (73.5%) contained phonemes that our preliminary phoneme study suggested might be illusionable. For these 147 words, we paired audio clips spoken by the speaker with illusory video dubs of the speaker saying the words with appropriately switched out phonemes. We tested these videos on 20 naive test subjects who did not participate in the preliminary study. Each subject watched half of the words and listened without video to the other half of the words, and were given the instructions: "Write down what you hear. What you hear may or may not be sensical. Also write down if a clip sounds unclear to you. Not that a clip may sound nonsensical but clear." Subjects were allowed up to three plays of each clip.
We found that watching the illusory videos led to an average miscomprehension rate of 24.8%, a relative 148% increase from the baseline of listening to the audio alone (Table 3). The illusory videos made people less confident about their correct answers, with an additional 5.1% of words being heard correctly but unclearly, compared to 2.1% for audio only. For 17% of the 200 words, the illusory videos increased the error rates by more than 30% above the audio-only baseline.
|Audio Only||Illusory Video||Relative Increase|
|Words Incorrectly Identified||10.0||24.8||+148%|
|Words Correct with Low Confidence||2.1||5.1||+144%|
3.2.1 Prediction model for word-level illusionability
To create a predictive model for word-level illusionability, we used illusionable phonemes enriched with positional information as features. Explicitly, for each of the 10 illusionable phonemes, we created three features from the phoneme being in an initial position (being the first phoneme of the word), a medial position (having phonemes come before and after), or a final position (being the last phoneme of the word). We then represented each word with a binary bag-of-words model Joachims (1998)
, giving each of the 30 phonemes-with-phonetic-context features a value of 1 if present in the word and 0 otherwise. We performed ridge regression on these features with a constant term. We searched for the optimalregularization constant among the values [0.1, 1, 10, 100] and picked the optimal one based on training set performance. The train:test split was in the proportion 85%:15% and was randomly chosen for each trial. Across 10k randomized trials, we obtain average training and test set correlations of and respectively.
Our final model achieves an out-of-sample correlation of 57% between predicted and observed illusionabilites. Here, the observed illusionability of the words is calculated as the difference between the accuracy of watchers of the illusory videos and the accuracy of listeners, where “accuracy” is defined as the fraction of respondents who were correct. For each word, the predicted illusionability is calculated from doing inference on that word using the averaged regression coefficients of the regression trials where the word is not in the training set.
Our predicted illusionability is also calibrated, in the sense that for the words predicted to have an illusionability <0.1, the mean empirical illusionability is 0.04; for words with predicted illusionability in the interval [0.1, 0.2] the mean empirical illusionability is 0.14; for predicted illusionability between [0.2, 0.3] the mean observed is 0.27; and for predicted illusionability >0.3, the mean observed is 0.50. Figure 2 visually depicts the match between the observed and predicted word illusionabilities.
3.3 Sentence-level experiments
We set up the following experiment on naturally occurring sentences. We randomly sampled 300 sentences of lengths 4-8 words inclusive from the novel Little Women (Alcott, 1994) from the Project Gutenberg corpus. From this reduced sample, we selected and perturbed 32 sentences that we expected to be illusionable (listed in Table 4). With the speaker, we prepared two formats of each sentence: original video (with original audio), and illusory video (with original audio).We then evaluated the perception of these on 1306 naive test subjects on Amazon Mechanical Turk.
The Turkers were shown videos for six randomly selected sentences, three illusory and three original, and were given the prompt: "Press any key to begin video [index #] of 6. Watch the whole video, and then you will be prompted to write down what the speaker said." Only one viewing of any clip was allowed, to simulate the natural setting of observing a live audio/video stream. Each Turker was limited to six videos to reduce respondent fatigue. Turkers were also asked to report their level of confidence in what they heard on a scale from no uncertainty (0%) to complete uncertainty (100%). One hundred and twelve Turkers (8.6%) did not adhere to the prompt, writing unrelated responses, and their results were omitted from analysis.
We found that watching the illusory videos led to an average miscomprehension rate of 32.8%, a relative 145% increase from the baseline of listening to the audio alone (Table 5). The illusory videos made people less confident about their correct answers. Turkers who correctly identified the audio message in an illusory video self-reported an average uncertainty of 42.9%, which is a relative 123% higher than the average reported by the Turkers who correctly understood the original videos. Examples of mistakes made by listeners of the illusory videos are shown in Table 6. Overall we found that for 11.5% of the 200 sampled sentences (23 out of the 30 videos we created), the illusory videos increased the error rates by more than 10% above the audio-only baseline.
|Illusionable||I’m not too proud to beg for Father.||No need of that.|
|I’d like to wear them, Mother can I?||It’s no harm, Jo|
|Now do be still, and stop bothering.||Well, I like that!|
|How many did you have out?||Open the door, Amy!|
|Nonsense, that’s of no use.||I am glad of that!|
|Of course that settled it.||I can’t bear saints.|
|Serves me right for trying to be fine.||Of course I am!|
|There’s gratitude for you!||That’s my good girl.|
|Capital boys, aren’t they?||You’ve got me, anyhow.|
|Brown, that is, sometimes.||I’ll tell you some day.|
|What do you know about him?||He won’t do that.|
|The plan of going over was not forgotten.|
|Not Illusionable||On the way get these things.||I’m glad of it.|
|Aren’t you going with him?||Then I’ll tell you.|
|That was the question.||That’s why I do it.|
I hate to have odd gloves!
|We don’t know him!|
|What good times they had, to be sure|
|Original Video||Illusory Video||Relative Increase|
|Sentences Incorrectly Identified||13.4%||32.8%||+145%|
|Uncertainty of Correct Sentences||19.4%||42.9%||+121%|
|Sentence||Sample Listener Perceptions|
|Serves me right for trying to be fine.||Serves thee right for trying to be thine|
|Serve you right for trying to be kind|
|serbes ye right shine by rhine|
|Of course that settled it.||up course bat saddle it|
|Of course Max settled it.|
|I can’t bear saints.||I can’t wear skates|
|I can’t hear saints.|
|I can’t spare sink|
|How many did you have out?||How many did you knock out?|
|How many did you help out|
|I’ll tell you some day.||I’ll tell you Sunday.|
|Ill tell you something|
|I’m not too proud to beg for Father.||I’m not too proud to beg or bother|
|I want the crowd to beg for father|
|Ine not too proud to beg for father|
|Well, I like that!||Well I like pets|
|Well, I like baths|
|Now do be still, and stop bothering.||now do we still end stop bothering|
|And do we still end sauce watering?|
|Of course I am!||Of course I an|
|Of course I can|
|Of course I anth.|
|You’ve got me, anyhow.||you got knee anyhow|
|You got the anyhow.|
|There’s gratitude for you!||Bears gratitude for you.|
|Pairs gratitudes or you.|
|Bears gratitude lore you|
3.3.1 Comparing word-level and sentence-level illusionabilities
We obtained a sentence-level illusionability prediction model with an out-of-sample correlation of 33% between predicted and observed illusionabilities. Here, the observed illusionability of the sentences was calculated as the difference between the accuracy of watchers of the illusory videos and the accuracy of watchers of original videos, where “accuracy” is defined as the fraction of respondents who were correct. We obtained predicted illusionabilities by simply using the maximum word illusionability prediction amongst the words in each sentence, with word predictions obtained from the word-level model. We attempted to improve our sentence-level predictive model by incorporating how likely the words appear under a natural language distribution, considering three classes of words: words in the sentence for which no illusion was attempted, the words for which an illusion was attempted, and the potentially perceived words for words for which an illusion was attempted. We used log word frequencies obtained from the the top 36.7k most common words from the Project Gutenberg corpus.222https://en.wiktionary.org/wiki/Wiktionary:Frequency_lists#Project_Gutenberg This approach could not attain better out-of-sample correlations than the naive method. This implies that context is important for sentence-level illusionability, and more complex language models should be used.
Finally, comparing word-level and sentence-level McGurk illusionabilities in natural speech, we observe that that the former is significantly higher. A greater McGurk effect at the word level is to be expected–sentences provide context with which the viewer could fill in confusions and misunderstandings. Furthermore, when watching a sentence video compared to a short word video, the viewer’s attention is more likely to stray, both from the visual component of the video, which evidently reduces the McGurk effect, as well as the from the audio component, which likely prompts the viewer to rely even more heavily on context. Nevertheless, there remains a significant amount of illusionability at the sentence-level.
4 Future Directions
This work is an initial step towards exploring the density of illusionable phenomena for humans. There are many natural directions for future work. In the vein of further understanding McGurk-style illusions, it seems worth building more accurate predictive models for sentence-level effects, and further investigating the security risks posed by McGurk illusions. For example, one concrete next step in understanding McGurk-style illusions would be to actually implement a system which takes an audio input, and outputs a video dub resulting in significant misunderstanding. Such a system would need to combine a high-quality speech-to-video-synthesis system (Suwajanakorn et al., 2017; Zakharov et al., 2019), with a fleshed-out language model and McGurk prediction model.
There is also the question of how to guard against “attacks” on human perception. For example, in the case of the McGurk effect, how can one rephrase a passage of text in such a way that the meaning is unchanged, but the rephrased text is significantly more robust to McGurk style manipulations? The central question in this direction is what fraction of natural language can be made robust without significantly changing the semantics.
A better understanding of when and why certain human perception systems are nonrobust can also be applied to make ML systems more robust. In particular, neural networks have been found to be susceptible to adversarial examples in automatic speech recognition(Schönherr et al., 2018; Carlini and Wagner, 2018) and to the McGurk effect (Ngiam et al., 2011), and a rudimentary approach to making language robust to the latter problem would be to use a reduced vocabulary that avoids words that score highly in our word-level illusionability prediction model. Relatedly, at the interface of cognitive science and adversarial examples, there has been work suggesting that humans can anticipate when or how machines will misclassify, including for adversarial examples (Chandrasekaran et al., 2017; Harding et al., 2018; Zhou and Firestone, 2019).
More broadly, as the tools for probing the weaknesses of ML systems develop further, it seems like a natural time to reexamine the supposed robustness of human perception. We anticipate unexpected findings. To provide one example, we summarize some preliminary results on audio-only illusions.
4.1 Auditory Illusions
An audio clip of the word “Laurel" gained widespread attention in 2018, with coverage by notable news outlets such as The New York Times and Time. Roughly half of listeners perceive “Laurel” and the other half perceive “Yanny” or similar-sounding words, with high confidence on both sides (Pressnitzer et al., 2018). One of the reasons the public was intrigued is because examples of such phenomena are viewed as rare, isolated instances. In a preliminary attempt to investigate the density of such phenomena, we identified five additional distinct examples (Table 7). The supplementary files include 10 versions of one of these examples, where listeners tend to perceive either “worlds” or “yikes.” Across the different audio clips, one should be able to hear both interpretations. The threshold for switching from one interpretation to another differs from person to person.
These examples were generated by examining 5000 words, and selecting the 50 whose spectrograms contain a balance of high and low frequency components that most closely matched those for the word “Laurel". Each audio file corresponded to the Google Cloud Text-to-Speech API synthesis of a word, after low frequencies were damped and the audio was slowed -x. After listening to these top 50 candidates, we evaluated the most promising five on a set of 15 individuals (3 female, 12 male, age range 22-33). We found multiple distributional modes of perceptions for all five audio clips. For example, a clip of “worlds” with the high frequencies damped and slowed down 1.5x was perceived by five listeners as “worlds”, four as “yikes/yites” and six as “nights/lights”. While these experiments do not demonstrate a density of such examples with respect to the set of all words—and it is unlikely that illusory audio tracks in this style can be created for the majority of words—they illustrate that even the surprising Yanny or Laurel phenomenon is not an isolated occurrence. It remains to be seen how dense such phenomena can be, given the right sort of subtle audio manipulation. ‘
Our work suggests that for a significant fraction of natural speech, human perception can be altered by using subtle, learnable perturbations. This is an initial step towards exploring the density of illusionable phenomenon for humans, and examining the extent to which human perception may be vulnerable to security risks like those that adversarial examples present for ML systems.
We hope our work inspires future investigations into the discovery, generation, and quantification of multimodal and unimodal audiovisual and auditory illusions for humans. There exist many open research questions on when and why humans are susceptible to various types of illusions, how to model the illusionability of natural language, and how natural language can be made more robust to illusory perturbations. Additionally, we hope such investigations inform our interpretations of the strengths and weaknesses of current ML systems. Finally, there is the possibility that some vulnerability to carefully crafted adversarial examples may be inherent to all complex learning systems that interact with high-dimensional inputs in an environment with limited data; any thorough investigation of this question must also probe the human cognitive system.
This research was supported in part by NSF award AF:1813049, an ONR Young Investigator award (N00014-18-1-2295), and a grant from Stanford’s Institute for Human-Centered AI. The authors would like to thank Jean Betterton, Shivam Garg, Noah Goodman, Kelvin Guu, Michelle Lee, Percy Liang, Aleksander Makelov, Jacob Plachta, Jacob Steinhardt, Jim Terry, and Alexander Whatley for useful feedback on the work. The research was done under Stanford IRB Protocol 46430.
- Alcott (1994) Louisa May Alcott. Little Women (1868-9). Roberts Brothers, 1994.
- Bastien-Toniazzo et al. (2010) Mireille Bastien-Toniazzo, Aurelie Stroumza, and Christian Cave. Audio-visual perception and integration in developmental dyslexia: An exploratory study using the mcgurk effect. Current psychology letters. Behaviour, brain & cognition, 25(3), 2010.
Behzadan and Munir (2017)
Vahid Behzadan and Arslan Munir.
Vulnerability of deep reinforcement learning to policy induction attacks.In
International Conference on Machine Learning and Data Mining in Pattern Recognition, pages 262–275. Springer, 2017.
- (4) Julien Besle, Alexandra Fort, Claude Delpuech, and Marie-H Giard. Bimodal speech: early suppressive visual effects in human auditory cortex.
- Carlini and Wagner (2018) Nicholas Carlini and David Wagner. Audio adversarial examples: Targeted attacks on speech-to-text. In 2018 IEEE Security and Privacy Workshops (SPW), pages 1–7. IEEE, 2018.
- Chandrasekaran et al. (2017) Arjun Chandrasekaran, Deshraj Yadav, Prithvijit Chattopadhyay, Viraj Prabhu, and Devi Parikh. It takes two to tango: Towards theory of ai’s mind. arXiv preprint arXiv:1704.00717, 2017.
Chen et al. (2018)
Pin-Yu Chen, Yash Sharma, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh.
Ead: elastic-net attacks to deep neural networks via adversarial
Thirty-second AAAI conference on artificial intelligence, 2018.
- Delbeuck et al. (2007) Xavier Delbeuck, Fabienne Collette, and Martial Van der Linden. Is alzheimer’s disease a disconnection syndrome?: Evidence from a crossmodal audio-visual illusory experiment. Neuropsychologia, 45(14):3315–3323, 2007.
- Elsayed et al. (2018) Gamaleldin F Elsayed, Shreya Shankar, Brian Cheung, Nicolas Papernot, Alex Kurakin, Ian Goodfellow, and Jascha Sohl-Dickstein. Adversarial examples that fool both human and computer vision. Advances in Neural Information Processing Systems, 2018.
- Evtimov et al. (2018) Ivan Evtimov, Kevin Eykholt, Earlence Fernandes, Tadayoshi Kohno, Bo Li, Atul Prakash, Amir Rahmati, and Dawn Song. Robust physical-world attacks on machine learning models. IEEE Conference on Computer Vision and Pattern Recognition, 2018.
- Harding et al. (2018) Samuel M Harding, Prashanth Rajivan, Bennett I Bertenthal, and Cleotilde Gonzalez. Human decisions on targeted and non-targeted adversarial samples. Proc. of the 40th Annual Conference of the Cognitive Science Society, pages 451–456, 2018.
- Hillis et al. (2002) James M Hillis, Marc O Ernst, Martin S Banks, and Michael S Landy. Combining sensory information: mandatory fusion within, but not between, senses. Science, 298(5598):1627–1630, 2002.
- Huang et al. (2017) Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies. International Conference on Learning Representations, 2017.
- Irwin et al. (2006) Julia R Irwin, DH Whalen, and Carol A Fowler. A sex difference in visual influence on heard speech. Perception & Psychophysics, 68(4):582–592, 2006.
Text categorization with support vector machines: Learning with many relevant features.In European conference on machine learning, pages 137–142. Springer, 1998.
- Kurakin et al. (2016) Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533, 2016.
- McGurk and MacDonald (1976) Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 264(5588):746, 1976.
- Mongillo et al. (2008) Elizabeth A Mongillo, Julia R Irwin, DH Whalen, Cheryl Klaiman, Alice S Carter, and Robert T Schultz. Audiovisual processing in children with and without autism spectrum disorders. Journal of autism and developmental disorders, 38(7):1349–1358, 2008.
Ngiam et al. (2011)
Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and Andrew Y
Multimodal deep learning.In Proceedings of the 28th international conference on machine learning, pages 689–696, 2011.
- Norrix et al. (2006) Linda W Norrix, Elena Plante, and Rebecca Vance. Auditory–visual speech integration by adults with and without language-learning disabilities. Journal of Communication Disorders, 39(1):22–36, 2006.
- Norrix et al. (2007) Linda W Norrix, Elena Plante, Rebecca Vance, and Carol A Boliek. Auditory-visual integration for speech by children with and without specific language impairment. Journal of Speech, Language, and Hearing Research, 50(6):1639–1651, 2007.
- Papernot et al. (2016) Nicolas Papernot, Patrick McDaniel, Arunesh Sinha, and Michael Wellman. Towards the science of security and privacy in machine learning. arXiv preprint arXiv:1611.03814, 2016.
- Papernot et al. (2017) Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia Conference on Computer and Communications Security, pages 506–519, 2017.
- Pressnitzer et al. (2018) Daniel Pressnitzer, Jackson Graves, Claire Chambers, Vincent De Gardelle, and Paul Egré. Auditory perception: Laurel and yanny together at last. Current Biology, 28(13):R739–R741, 2018.
- Samhita and Gross (2013) Laasya Samhita and Hans J Gross. The “clever hans phenomenon” revisited. Communicative & integrative biology, 6(6):e27122, 2013.
- Schönherr et al. (2018) Lea Schönherr, Katharina Kohls, Steffen Zeiler, Thorsten Holz, and Dorothea Kolossa. Adversarial attacks against automatic speech recognition systems via psychoacoustic hiding. arXiv preprint arXiv:1808.05665, 2018.
- Sekiyama (1997) Kaoru Sekiyama. Cultural and linguistic factors in audiovisual speech processing: The mcgurk effect in chinese subjects. Perception & Psychophysics, 59(1):73–80, 1997.
- Sekiyama et al. (2014) Kaoru Sekiyama, Takahiro Soshi, and Shinichi Sakamoto. Enhanced audiovisual integration with aging in speech perception: a heightened mcgurk effect in older adults. Frontiers in Psychology, 5:323, 2014.
- Summerfield (1992) Quentin Summerfield. Lipreading and audio-visual speech perception. Phil. Trans. R. Soc. Lond. B, 335(1273):71–78, 1992.
- Suwajanakorn et al. (2017) Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics, 36(4):95, 2017.
- Youse et al. (2004) Kathleen M Youse, Kathleen M Cienkowski, and Carl A Coelho. Auditory-visual speech perception in an adult with aphasia. Brain injury, 18(8):825–834, 2004.
- Zakharov et al. (2019) Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. arXiv preprint arXiv:1905.08233, 2019.
- Zhou and Firestone (2019) Zhenglong Zhou and Chaz Firestone. Humans can decipher adversarial images. Nature communications, 10(1):1334, 2019.