Both automatic speech recognition and text to speech systems need accurate pronunciations, typically obtained by using both a lexicon dictionary and a grapheme to phoneme (G2P) model. G2Ps typically struggle with predicting pronunciations for tail words, and we hypothesized that one reason is because they try to discover general pronunciation rules without using prior knowledge of the pronunciation of related words. Our new approach expands a sequence-to-sequence G2P model by injecting prior knowledge. In addition, our model can be updated without having to retrain a system. We show that our new model has significantly better performance for German, both on a tightly controlled task and on our real-world system. Finally, the simplification of the system allows for faster and easier scaling to other languages.

No abstract

Prediction of morphological forms is a well-studied problem and can lead to better speech systems either directly by rescoring models for correcting morphology, or indirectly by more accurate dialog systems with improved natural language generation and understanding. This includes both lemmatization, i.e. deriving the lemma or root word from a given surface form as well as morphological inflection, i.e. deriving surface forms from the lemma. We train and evaluate various languageagnostic end-to-end neural sequence-to-sequence models for these tasks and compare their effectiveness. We further augment our models with pronunciation information which is typically available in speech systems to further improve the accuracies of the same tasks. We present the results across both morphologically modest and rich languages to show robustness of our approach.

When scaling to low resource languages for speech synthesis or speech recognition in an industrial setting, a common challenge is the absence of a readily available pronunciation lexicon. Common alternatives are handwritten letter-to-sound rules and data-driven grapheme-to-phoneme (G2P) models, but without a pronunciation lexicon it is hard to even determine their quality. We identify properties of a good quality metric and note drawbacks of naïve estimates of G2P quality in the domain of small test sets. We demonstrate a novel method for reliable evaluation of G2P accuracy with minimal human effort. We also compare behavior of known state-of-the-art approaches for training with limited data. Finally we evaluate a new active learning approach for training G2P models in the low resource setting.

Data-driven algorithm configuration [16,1] is a promising, learning-based approach for beyond worst-case analysis of algorithms with tunable parameters. An important open problem is the design of efficient data-driven algorithms for algorithm families with more than one parameter. In this work we provide algorithms for efficient (output-polynomial) multidimensional parameter tuning, i.e. for families with a small constant number of parameters, for three very different combinatorial problems -linkage-based clustering, dynamic programming for sequence alignment, and auction design for two-part tariff schemes. We extend the single-parameter clustering algorithm of [4] to multiple parameters and to the sequence alignment problem by proposing an execution graph which compactly represents all the states the algorithm could attain for all possible parameter values. A key problem-specific challenge is to efficiently compute how the partition of the parameter space (into regions with unique algorithmic states) changes with a single algorithmic step. We give algorithms which improve on the runtime of previously best known results for linkage-based clustering, sequence alignment and two-part tariff pricing.