For an automatic speech recognition system to produce sensibly formatted, readable output, the spoken-form token sequence produced by the core speech recognizer must be converted to a written-form string. This process is known as inverse text normalization (ITN). Here we present a mostly data-driven ITN system that leverages a set of simple rules and a few handcrafted grammars to cast ITN as a labeling problem. To this labeling problem, we apply a compact bi-directional LSTM. We show that the approach performs well using practical amounts of training data.

No abstract

When deploying a spoken dialogue system in a new domain, one faces a situation where little to no data is available to train domain-specific statistical models. We describe our experience with bootstrapping a dialogue system for public transit and weather information in real-word deployment under public use. We proceeded incrementally, starting from a minimal system put on a toll-free telephone number to collect speech data. We were able to incorporate statistical modules trained on collected data -in-domain speech recognition language models and spoken language understanding -while simultaneously extending the domain, making use of automatically generated semantic annotation. Our approach shows that a successful system can be built with minimal effort and no in-domain data at hand.

This paper presents an extension of the Kaldi automatic speech recognition toolkit to support on-line recognition. The resulting recogniser supports acoustic models trained using state-of-theart acoustic modelling techniques. As the recogniser produces word posterior lattices, it is particularly useful in statistical dialogue systems, which try to exploit uncertainty in the recogniser's output. Our experiments show that the online recogniser performs significantly better in terms of latency when compared to a cloud-based recogniser.

No abstract