Towards Improving Low-Resource Speech Recognition Using Articulatory and Language Features

Towards Improving Low-Resource Speech Recognition Using Articulatory and Language Features Markus M¨uller, Sebastian St¨uker, Alex Waibel Institute for Anthropomatics, Karlsruhe Institute of Technology, Karlsruhe, Germany {m.mueller|sebastian.stueker|waibel}@kit.edu Abstract In an increasingly globalized world, there is a rising de- mand for speech recognition systems. Systems for languages like English, German or French do achieve a decent perfor- mance, but there exists a long tail of languages for which such systems do not yet exist. State-of-the-art speech recog- nition systems feature Deep Neural Networks (DNNs). Be- ing a data driven method and therefore highly dependent on sufficient training data, the lack of resources directly af- fects the recognition performance. There exist multiple tech- niques to deal with such resource constraint conditions, one approach is the use of additional data from other languages. In the past, is was demonstrated that multilingually trained systems benefit from adding language feature vec- tors (LFVs) to the input features, similar to i-Vectors. In this work, we extend this approach by the addition of articulatory features (AFs). We show that AFs also benefit from LFVs and that multilingual system setups benefit from adding both AFs and LFVs. Pretending English to be a low-resource lan- guage, we restricted ourselves to use only 10h of English acoustic training data. For system training, we use additional data from French, German and Turkish. By using a combi- nation of AFs and LFVs, we were able to decrease the WER from 18.1% to 17.3% after system combination in our setup using a multilingual phone set. 1. Introduction Language and speech technologies have matured dramati- cally in recent years. With the emergence of these technolo- gies into our daily lives and an increasingly globalized world, there is a growing demand for developing systems for new languages. Since many methods for system building are data- driven, a certain amount of training data is required. While these resources exist for languages like English, only a few other languages are as well researched and have a compara- ble amount of training data readily available. There are approximately 7000 living languages in the world [1], many of them facing extinction. With the majority not being of social or economic interest, special methods are required to handle conditions like, e.g., sparseness of data. A common approach to build a speech recognition systems in resource constrained conditions is to incorporate data from well-resourced, data-rich languages. In the past, it has been shown that using data from multiple languages is beneficial in cases where only a limited amount of training data from the target language is available. Up until now, we only considered neural networks trained using phonemes as targets. We trained networks using multi- lingual phone sets, covering the phoneme inventory of multi- ple languages. The phoneme inventory, even if based on mul- tiple languages, is limited as phonemes represent a certain configuration of the articulators in the vocal tract. A limited phone set can therefore only represent a limited amount of different configurations. Considering each articulatory fea- ture like the …

Content truncated to 500 words