We describe a novel way to implement subword language models in speech recognition systems based on weighted finite state transducers, hidden Markov models, and deep neural networks. The acoustic models are built on graphemes in a way that no pronunciation dictionaries are needed, and they can be used together with any type of subword language model, including character models. The advantages of short subword units are good lexical coverage, reduced data sparsity, and avoiding vocabulary mismatches in adaptation. Moreover, constructing neural network language models (NNLMs) is more practical, because the input and output layers are small. We also propose methods for combining the benefits of different types of language model units by reconstructing and combining the recognition lattices. We present an extensive evaluation of various subword units on speech datasets of four languages: Finnish, Swedish, Arabic, and English. The results show that the benefits of short subwords are even more consistent with NNLMs than with traditional n-gram language models. Combination across different acoustic models and language models with various units improve the results further. For all the four datasets we obtain the best results published so far. Our approach performs well even for English, where the phoneme-based acoustic models and word-based language models typically dominate: The phoneme-based baseline performance can be reached and improved by 4% using graphemes only when several grapheme-based models are combined. Furthermore, combining both grapheme and phoneme models yields the state-of-the-art error rate of 15.9% for the MGB 2018 dev17b test. For all four languages we also show that the language models perform reasonably well when only limited training data is available.

We describe the speech recognition systems we have created for MGB-3, the 3rd Multi Genre Broadcast challenge, which this year consisted of a task of building a system for transcribing Egyptian Dialect Arabic speech, using a big audio corpus of primarily Modern Standard Arabic speech and only a small amount (5 hours) of Egyptian adaptation data. Our system, which was a combination of different acoustic models, language models and lexical units, achieved a Multi-Reference Word Error Rate of 29.25%, which was the lowest in the competition. Also on the old MGB-2 task, which was run again to indicate progress, we achieved the lowest error rate: 13.2%.

The result is a combination of the application of state-of-the-art speech recognition methods such as simple dialect adaptation for a Time-Delay Neural Network (TDNN) acoustic model (-27% errors compared to the baseline), Recurrent Neural Network Language Model (RNNLM) rescoring (an additional -5%), and system combination with Minimum Bayes Risk (MBR) decoding (yet another -10%). We also explored the use of morph and character language models, which was particularly beneficial in providing a rich pool of systems for the MBR decoding.

Today, the vocabulary size for language models in large vocabulary speech recognition is typically several hundreds of thousands of words. While this is already sufficient in some applications, the out-of-vocabulary words are still limiting the usability in others. In agglutinative languages the vocabulary for conversational speech should include millions of word forms to cover the spelling variations due to colloquial pronunciations, in addition to the word compounding and inflections. Very large vocabularies are also needed, for example, when the recognition of rare proper names is important.

Because in agglutinative languages the number of observed word forms is very high, subword units are often utilized in speech recognition. However, the proper use of subword units requires careful consideration of details such as silence modeling, position-dependent phones, and combination of the units. In this paper, we implement subword modeling in the Kaldi toolkit by creating modified lexicon by finite-state transducers to represent the subword units correctly. We experiment with multiple types of word boundary markers and achieve the best results by adding a marker to the left or right side of a subword unit whenever it is not preceded or followed by a word boundary, respectively. We also compare three different toolkits that provide data-driven subword segmentations. In our experiments on a variety of Finnish and Estonian datasets, the best subword models do outperform word-based models and naive subword implementations. The largest relative reduction in WER is a 23% over word-based models for a Finnish read speech dataset. The results are also better than any previously published ones for the same datasets, and the improvement on all datasets is more than 5%.