Minimum Phone Error model training on merged acoustic units for transcribing bilingual code-switched speech

Abstract

This paper proposes to perform Minimum Phone Error (MPE) model training on merged acoustic units for transcribing Mandarin-English code-switched lectures with highly imbalanced language distribution. Some of the acoustic events in Mandarin and English may have very similar characteristics, so the states or Gaussian mixtures representing them can be merged with identical shared parameters. When MPE is performed afterwards, these merged identical states or Gaussian mixtures can form a compact acoustic unit set. In this way MPE can better discriminate the acoustic units of both languages, because similar units are merged while distinct units are differentiated. Significant improvements in recognition accuracy were observed in the preliminary experiments on real-world bilingual code-switched lecture corpus recorded at National Taiwan University.

Synthesising isiZulu-English Code-Switch Bigrams Using Word Embeddings

“…Small corpora have been compiled for English-Spanish [1,2], Cantonese-English [3,4], Hindi-English [5] and for Sepedi-English [6]. However, the language pair English-Mandarin has received by far the most attention [7][8][9][10][11][12][13][14]. Approaches to code-switched language modelling include interpolating n-gram language models (LM) trained on monolingual data [13], n-grams trained on code-switched data [5,7], class-based n-grams using additional features [4], recurrent neural networks [10], and combinations of approaches [11].…”

“…However, the language pair English-Mandarin has received by far the most attention [7][8][9][10][11][12][13][14]. Approaches to code-switched language modelling include interpolating n-gram language models (LM) trained on monolingual data [13], n-grams trained on code-switched data [5,7], class-based n-grams using additional features [4], recurrent neural networks [10], and combinations of approaches [11]. A particularly relevant recent study considered features for factored language models for Mandarin-English code-switched speech using the SEAME corpus [12].…”