Speaker: Sergio Segovia. Abstract: The core idea is to predict latent representations of the full input data based on a masked view of the input in a self-distillation setup using a standard Transformer architecture. Instead of predicting modality-specific targets such… Read More
Data Augmentation for Decoupled Calibration of Deep Neural Network Classifiers
Speaker: Sergio Márquez Carrero. Abstract: Modern Deep Neural Networks (DNN) have significantly outperformed those employed over a decade ago in terms of accuracy. Nonetheless, the outputs generated by these models are poorly calibrated, causing substantial issues in a variety of… Read More
Connectionist Temporal Classification (CTC) Speech Segmentation
Speaker: W. Fernando López Gavilanez. Abstract: Motivated by the lack of high-quality labeled data for specific scenarios, such as emergencies in the home environment, we explored a CTC-segmentation method to generate a specific-purpose speech dataset. The project seeks the quality improvement of… Read More
BigSSL: Large-Scale Semi-Supervised Learning for ASR
Speaker: Laura Herrera Abstract: This paper deals with results obtained on very large automatic speaker recognition models.A large amount of labelled data is not always available and sometimes they do not generalize enough. Consequently, the authors propose to use pre-trained… Read More
Efficient Neural Approaches for Automatic Speech Recognition
Speaker: Doroteo Torre Toledano Abstract: Many different end-to-end neural approaches have been proposed in the last years in the field of automatic speech recognition (ASR). However, most of the research available compares systems only in terms of accuracy (word error… Read More
Structured Output Learning
Speaker: María Pilar Fernández Rodríguez Abstract: Speech applications dealing with conversations require not only recognizing the spoken words, but also determining who spoke when, the language, punctuation, capitalization… To deal with it, it is typically addressed by merging the outputs… Read More
Voxceleb Experiment: fairness
Speaker: Almudena Aguilera Abstract: The experiment is based on the dataset from Voxceleb [1], using the two pre-trained models. The main idea of these experiments was to study the fairness problems in different demographic groups present in the data base… Read More
Semi-Supervised Music Tagging Transformer
Speaker: David Martín Abstract: Music Tagging Transformer (MTT) was recently released in the latest ISMIR 2021 Conference as one of the most erupting deep learning approaches for Music Information Retrieval. It consists of a semi-supervised approach where the model captures… Read More
Encoder-Decoder Based Attractor Calculation for End-to-End Neural Diarization
Speaker: Alicia Lozano Díez Abstract: In this talk, we will deeply review the algorithms behind end-to-end systems for speaker diarization based on neural networks. In particular, we will describe how the encoder-decoder part of the model calculates “attractors” that capture… Read More
Unsupervised Sound Separation Using Mixture Invariant Training
Speaker: Diego de Benito Gorrón Abstract: In recent years, rapid progress has been made on the problem of single-channel sound separation using supervised training of deep neural networks. In such supervised approaches, a model is trained to predict the component… Read More