Minimally Supervised Number Normalization

Abstract

We propose two models for verbalizing numbers, a key component in speech recognition and synthesis systems. The first model uses an end-to-end recurrent neural network. The second model, drawing inspiration from the linguistics literature, uses finite-state transducers constructed with a minimal amount of training data. While both models achieve near-perfect performance, the latter model can be trained using several orders of magnitude less data than the former, making it particularly useful for low-resource languages.

Weird Inflects but OK: Making Sense of Morphological Generation Errors

“…Silly errors This category consists of those "bizarre" errors which defy any purely linguistic characterization. In addition to the aforementioned case of *membled, such errors have also been reported for other language generation tasks such as machine translation (Arthur et al 2016) and text normalization (Gorman and Sproat 2016, Sproat and Jaitly 2017, Zhang et al 2019.…”

Unified Verbalization for Speech Recognition & Synthesis Across Languages

“…Our general process of minimally supervised number names induction is described in [4]. In essence, we use a set of training data consisting of digits mapped to their verbalization (like 123 → one hundred twenty three) to induce a finite state transducer (FST) which can produce the factorization for any number.…”

“…Verbalizers for most semiotic classes depend on underlying core number names grammars specifying the verbalization of numbers like English one, two, three. We first describe how we modify the induction algorithm in [4] to build these number names grammars across a wider range of languages. We then describe a system which builds on this algorithm to induce verbalization grammars for ASR and TTS systems alike.…”

“…This variation, and the recursive complexity of number naming in general, makes them less amenable to templates than other semiotic classes. [4] induce the basic factorization and other features of number names systems from a small set of labeled examples. In Section 2, we describe some extensions to this system which provide better support for factorizations commonly found in the languages of South Asia and sub-Saharan Africa.…”

Dual Encoder Classifier Models as Constraints in Neural Text Normalization

“…One solution to this problem is to use covering grammars, (usually) finite-state models that can constrain the neural models to a reasonable (context-independent) space of options so that 2mA could be read as two milliamperes or two m a, but not two million liters [3]. These covering grammars can be learned in whole or in part from data [7].…”