This paper introduces FT SPEECH, a new speech corpus created from the recorded meetings of the Danish Parliament, otherwise known as the Folketing (FT). The corpus contains over 1,800 hours of transcribed speech by a total of 434 speakers. It is significantly larger in duration, vocabulary, and amount of spontaneous speech than the existing public speech corpora for Danish, which are largely limited to read-aloud and dictation data. We outline design considerations, including the preprocessing methods and the alignment procedure. To evaluate the quality of the corpus, we train automatic speech recognition systems (ASR) on the new resource and compare them to the systems trained on the Danish part of Språkbanken, the largest public ASR corpus for Danish to date. Our baseline results show that we achieve a 14.01 WER on the new corpus. A combination of FT SPEECH with in-domain language data provides comparable results to models trained specifically on Språkbanken, showing that FT SPEECH transfers well to this data set. Interestingly, our results demonstrate that the opposite is not the case. This shows that FT SPEECH provides a valuable resource for promoting research on Danish ASR with more spontaneous speech.

Phonological features can indicate word class and we can use word class information to disambiguate both homophones and homographs in automatic speech recognition (ASR). We show Danish stød can be predicted from speech and used to improve ASR. We discover which acoustic features contain the signal of stød, how to use these features to predict stød and how we can make use of stød and stødpredictive acoustic features to improve overall ASR accuracy and decoding speed. In the process, we discover acoustic features that are novel to the phonetic characterisation of stød.

Deep Neural Network (DNN) acoustic models are an essential component in automatic speech recognition (ASR). The main sources of accuracy improvements in ASR involve training DNN models that require large amounts of supervised data and computational resources. While the availability of sufficient monolingual data is a challenge for low-resource languages, the computational requirements for resource rich languages increases significantly with the availability of large data sets. In this work, we provide novel solutions for these two challenges in the context of training a feed-forward DNN acoustic model (AM) for mobile voice search. To address the datasparsity challenge, we bootstrap our multilingual AM using data from languages in the same language family. To reduce training time, we use cyclical learning rate (CLR) which has demonstrated fast convergence with competitive or better performance when training neural networks on tasks related to text and images. We reduce training time for our Mandarin Chinese AM with 81.4% token accuracy from 40 to 21.3 hours and increase the word accuracy on three romance languages by 2-5% with multilingual AMs compared to monolingual DNN baselines.