Deep learning for spoken language identification
Matias Lindgren
Abstract
This thesis applies deep learning based classification techniques to identify natural languages from speech. The primary motivation behind this thesis is to implement accurate techniques for segmenting multimedia materials by the languages spoken in them. \n \n \nSeveral existing state-of-the-art, deep learning based approaches are discussed and a subset of the discussed approaches are selected for quantitative experimentation. The selected model architectures are trained on several well-known spoken language identification datasets containing several different languages. Segmentation granularity varies between models, some supporting input audio lengths of 0.2 seconds, while others require 10 second long input to make a language decision. \n \n \nResults from the thesis experiments show that an unsupervised representation of acoustic units, produced by a deep sequence-to-sequence auto encoder, cannot reach the language identification performance of a supervised representation, produced by a multilingual phoneme recognizer. Contrary to most existing results, in this thesis, acoustic-phonetic language classifiers trained on labeled spectral representations outperform phonotactic classifiers trained on bottleneck features of a multilingual phoneme recognizer. More work is required, using transcribed datasets and automatic speech recognition techniques, to investigate why phoneme embeddings did not outperform simple, labeled spectral features. \n \n \nWhile an accurate online language segmentation tool for multimedia materials could not be constructed, the work completed in this thesis provides several insights for building feasible, modern spoken language identification systems. As a side-product of the experiments performed during this thesis, a free open source spoken language identification software library called "lidbox" was developed, allowing future experiments to begin where the experiments of this thesis end.