Exploiting temporal context in speech technologies using lstm recurrent neural networks
- Autores: Rubén Zazo Candil
- Directores de la Tesis: Joaquín González Rodríguez (dir. tes.), Javier González Domínguez (dir. tes.)
- Lectura: En la Universidad Autónoma de Madrid ( España ) en 2018
- Idioma: inglés
- Número de páginas: 120
- Tribunal Calificador de la Tesis: Luis Alfonso Hernández Gómez (presid.), Doroteo Torre Toledano (secret.), Ascensión Gallardo Antolín (voc.), Alfonso Ortega Giménez (voc.), Ricardo de Córdoba Herralde (voc.)
- Programa de doctorado: Programa de Doctorado en Ingeniería Informática y de Telecomunicación por la Universidad Autónoma de Madrid
- Calificación: Sobresaliente Cum Laude
- Materias:
- Enlaces
- Tesis en acceso abierto en: Biblos-e Archivo
- Resumen
The importance of speech technologies is today out of the question. Speech is the most natural way to communicate and speech technologies have reached a point where operating our devices with the voice is not science fiction anymore. Google’s assistant, Apple’s Siri or Amazon’s Alexa are the proof that we are already shifting towards a new paradigm that has the potential to change the way we interact with our devices in a similar way to the mouse, the icons or windows did a few years ago.
Among all speech related technologies, automatic speech recognition (ASR), devoted to producing a readable text transcription from an audio signal, is the most popular and one of the hottest topics in the machine learning community. This task, which is e↵ortless for humans, has taken decades to develop and most of the biggest research teams in this community -in companies and universities- are focused on improving this technology, pushing the state-of-theart at an astonishing speed. Nevertheless, much of the information contained in the speech signal is actually contained in the context, where context means particular dialog state, options presented to the user, conversation topic, location, etc. Furthermore, voice-operated systems need approaches to solve many related tasks such as voice activity detection (to deal with noise and reduce computation), language identification (to select the proper ASR model) or hotword recognition (to wake up the system).
If there is a technology that has changed completely the way researchers looked at speech recognition in the last years it is deep learning. Neural networks are not a new method to pattern recognition but it is in the last decade when they re-surged and, in a very short period of time, outperformed many well-established approaches in most of the main machine learning tasks. When they were first proposed, the lack of computational power and the scarcity of data made impossible to train properly these models. Nowadays, the availability of graphic process units (GPUs) and large-scale databases allow us to apply successfully these methods obtaining results that were completely out of reach not that long ago. In addition to this, deep learning has greatly improved in the last few years and new models have been proposed improving the learning capabilities of neural networks such as convolutional neural networks for image recognition or recurrent neural networks for sequence learning. Among all these models, a special kind of recurrent neural network called long short-term memory (LSTM) is showing breathtaking results in most of the problems where a sequence is present, including automatic speech recognition, as it is capable of modeling the temporal context inherent to these sequences extracting a lot more information than feed forward deep neural networks do.
In this context, the present PhD Thesis pretends to test the limits of the most successful algorithms in the tasks that gather most of the attention from big research groups to determine how well they generalize across problems and to adapt this architecture to other related tasks where temporal context can contain relevant information. Thus, the main goal of this Dissertation is to provide a good understanding on long short-term memory recurrent neural networks (LSTMs) and to propose novel approaches based on them to language identification, voice activity detection and age estimation, improving as a result the state-of-the-art in the mentioned fields.
The research work described in this Dissertation has led to novel contributions which include the development of: i) an approach to language identification based on LSTMs using modest resources, ii) a state-of-the-art voice activity detector capable of both feature extraction and classification developed in Google Inc. that is currently in production working on more than 20 languages and millions of phones, and iii) a novel end-to-end approach to age estimation capable of dealing with short utterances. Moreover, di↵erent original experimental studies have been carried out during the development of the Thesis. Besides, the research work completed throughout the Thesis has been complemented with the generation of several research publications, talks and a patent currently in production.
