Keyvan Mohajer

The automatic measurement of facial beauty

We develop an automatic facial beauty scoring system based on ratios between facial features. After isolating the face, eyes, eyebrows and mouth in a portrait photograph, we represent a face abstractly as an 8-element vector of ratios between these features. We use a variant of the K-nearest neighbor algorithm, in the context of a parameterized metric space optimized using a genetic algorithm, to learn a beauty assignment function from a training set of photographs rated by humans. We assess performance on a test set of photographs, concluding that when facial ratios are accurately extracted in the computer vision phase, the results of the program are highly correlated with median-human ratings of beauty.

Driver drowsiness detection using multimodal sensor fusion

This paper proposes a multi-modal sensor fusion algorithm for the estimation of driver drowsiness. Driver sleepiness is believed to be responsible for more than 30% of passenger car accidents and for 4% of all accident fatalities. In commercial vehicles, drowsiness is blamed for 58% of single truck accidents and 31% of commercial truck driver fatalities. This work proposes an innovative automatic sleep-onset detection system. Using multiple sensors, the driver’s body is studied as a mechanical structure of springs and dampeners. The sleep-detection system consists of highly sensitive triple-axial accelerometers to monitor the driver’s upper body in 3-D. The subject is modeled as a linear time-variant (LTV) system. An LMS adaptive filter estimation algorithm generates the transfer function (i.e. weight coefficients) for this LTV system. Separate coefficients are generated for the awake and asleep states of the subject. These coefficients are then used to train a neural network. Once trained, the neural network classifies the condition of the driver as either awake or asleep. The system has been tested on a total of 8 subjects. The tests were conducted on sleep-deprived individuals for the sleep state and on fully awake individuals for the awake state. When trained and tested on the same subject, the system detected sleep and awake states of the driver with a success rate of 95%. When the system was trained on three subjects and then retested on a fourth “unseen” subject, the classification rate dropped to 90%. Furthermore, it was attempted to correlate driver posture and sleepiness by observing how car vibrations propagate through a person’s body. Eight additional subjects were studied for this purpose. The results obtained in this experiment proved inconclusive which was attributed to significant differences in the individual habitual postures.

Estimation-pruning (EP) algorithm for point-to-point travel cost minimization in a non-FIFO dynamic network

This paper presents the estimation-pruning (EP) algorithm for finding the best path (with minimum cost) from a source to a destination in a dynamic network that does not necessarily obey the first-in-first-out (FIFO) property. The EP algorithm consists of two steps. The first step is the forward or the estimation step in which a bound on the traveling cost of each possible path is calculated. The second step is the backward or the pruning step in which the paths that are unlikely to produce the best route are eliminated. The resulting network is then expanded in time and is converted to a static network, which is used to find the best route.

CommonSense: a preprocessing system to identify errors in large transcribed corpora

A system was designed to locate and correct errors in large transcribed corpora. The program, called CommonSense, relies on a set of rules that identify mistakes related to homonyms, words with distinct definitions but identical pronunciations. The system was run on the 1996 and 1997 Broadcast News Speech Corpora, and correctly identified more than 400 errors in these data. Future work may extend CommonSense to automatically correct errors in hypothesis files created as the output of speech recognition systems.

Robust speech recognition using time boundary detection

This paper explores the benefits of including time boundary information in Hidden Markov Model based speech recognition systems. Traditional systems normally feed the parameterized data into the HMM recognizer, which result in relatively complicated models and computationally expensive search steps. We propose a few methods of detecting time boundaries prior to parameterization, and present a novel way of including this additional information in the recognizer. The result is significant simplification in the model prototypes, higher accuracy and faster performance.

Post Recognition Speech Localization

How much does knowledge regarding a certain spoken word or phrase help with its localization? This is a very fundamental question for speech processing, and will be partially addressed in this paper. In particular, this work will utilize prior information regarding the contents of a speech signal in order to improve the artificial localization of it using Time delay of arrival (TDOA) between two microphones. The prior information, which is used to develop a very simple frequency-selective phase transform (FPT), increases the effective SNR by only using a subset of the highest SNR frequencies in the Phase Transform. Simulations in a reverberant environment show that the proposed approach can more robustly and accurately localize speech sources. For 20 ms signal segments, it is shown that using a subset of 45 percent of available speech frequency bins is superior to using 30, 60, or 100, where using 100 corresponds to the standard Phase Transform.