The most outstanding articles on AI research are compiled here. It is a carefully curated list of the most recent developments in data science and artificial intelligence, presented chronologically with a link to a longer article for more information.

Learning Individual Speaking Styles for Accurate Lip-to-Speech Synthesis

Humans instinctively rely on lip movements to deduce parts of a conversation when the voice is missing or distorted by extraneous noise. The researchers in this study investigate lip-to-speech synthesis, which involves generating natural speech based solely on a speaker's lip motions. Recognizing the significance of contextual and speaker-specific cues in achieving precise lip-reading, they diverge from previous research. The researchers are concentrating on acquiring precise lip sequences to speech mappings for individual speakers in unrestricted, extensive vocabulary scenarios. They gather and publish a comprehensive benchmark dataset, the first of its type, designed for training and assessing the single-speaker lip-to-speech job in real-life environments. 

The researchers suggest a new method with important design decisions to accomplish precise, lifelike lip-to-speech synthesis in unrestricted situations for the first time. Their strategy is four times more comprehensible than prior efforts in this field based on a thorough assessment using quantitative and qualitative metrics and human evaluation.

Generalized Zero-Shot Learning via Over-Complete Distribution

An adept and versatile deep neural network (DNN) should exhibit resilience against familiar and unfamiliar classes. Most supervised deep neural network techniques experience a decline in performance when dealing with classes that were not present in the training set. The researchers suggest creating an Over-Complete Distribution (OCD) using a Conditional Variational Autoencoder (CVAE) for training a discriminative classifier that performs well in Zero-Shot Learning (ZSL) scenarios with both seen and unseen classes. To enhance class separability and minimize class dispersion, they suggest using Online Batch Triplet Loss (OBTL) and Center Loss (CL) on the generated OCD. Zero-Shot Learning and Generalized Zero-Shot Learning protocols assess the framework's effectiveness on three benchmark databases: SUN, CUB, and AWA2. The results indicate that creating over-complete distributions and compelling the classifier to acquire a transformation function from overlapping to non-overlapping distributions can enhance performance on familiar and unfamiliar classes.

Concept Drift Detection for Multivariate Data Streams and Temporal Segmentation of Daylong Egocentric Videos

Temporal segmentation is crucial for many higher-level inference tasks due to egocentric movies' lengthy and unstructured format. The wearer's activities in an egocentric film usually extend over hours and are frequently interspersed with slow, incremental alterations. Moreover, the alteration of camera perspective resulting from the wearer's head movement leads to frequent, drastic, yet false scene transitions. Traditional Markov Random Field (MRF) pipelines need to be more effective due to the continuous nature of boundaries. In contrast, deep Long Short Term Memory (LSTM) networks are limited in capturing context within a few hundred frames, making them ineffective for egocentric films.

The researchers introduce a new unsupervised temporal segmentation technique for day-long egocentric recordings. The challenge is identifying idea drift in a time-varying, non-independent and identically distributed sequence of frames. Statistically limited thresholds are computed to identify idea drift between two temporally adjacent multivariate data segments with distinct underlying distributions while ensuring control over false positives. The determined threshold, which reflects confidence in the prediction, can also be utilized to regulate the level of detail in the output segmentation.

Image source: Unsplash