Reducing development costs of large vocabulary speech recognition systems

Résumé :
One of the outstanding challenges in large vocabulary automatic speech recognition (ASR) is the reduction of development costs required to build a new recognition system or adapt an existing one to a new task, language or dialect. The state-of-the-art ASR systems are based on the principles of the statistical learning paradigm, using information provided by two stochastic models, an acoustic (AM) and a language (LM) model. The standard methods used to estimate the parameters of such models are founded on two main assumptions : the training data sets are large enough, and the training data match well the target task. It is well-known that a great part of system development costs is due to the construction of corpora that fulfill these requirements. In particular, manually transcribing the audio data is the most expensive and time-consuming endeavor. For some applications, such as the recognition of low resourced languages or dialects, finding and collecting data is also a hard
(and expensive) task. As a means to lower the cost required for ASR system development, this thesis proposes and studies methods that aim to alleviate the need for manually transcribing audio data for a given target task. Two axes of research are explored. First, unsupervised training methods are explored in order to build three of the main components of ASR systems : the acoustic model, the multi-layer perceptron (MLP) used to extract acoustic features and the language model. The unsupervised training methods aim to estimate the model parameters using a large amount of automatically (and inaccurately) transcribed audio data, obtained thanks to an existing recognition system. A novel method for unsupervised AM training that copes well with the automatic audio transcripts is proposed : the use of multiple recognition hypotheses (rather than the best one) leads to consistent gains in performance over the standard approach. Unsupervised MLP training is proposed as an alternati
ve to build efficient acoustic models in a fully unsupervised way. Compared to cross-lingual MLPs trained in a supervised manner, the unsupervised MLP leads to competitive performance levels even if trained on only about half of the data amount. Unsupervised LM training approaches are proposed to estimate standard back-off n-gram and neural network language models. It is shown that unsupervised LM training leads to additive gains in performance on top of unsupervised AM training. Second, this thesis proposes the use of model interpolation as a rapid and flexible way to build task specific acoustic models. In reported experiments, models obtained via interpolation outperform the baseline pooled models and equivalent maximum a posteriori (MAP) adapted models. Interpolation proves to be especially useful for low resourced dialect ASR. When only a few (2 to 3 hours) or no acoustic data truly matching the target dialect are available for AM training, model interpolation leads to
substantial performance gains compared to the standard training methods.