Context-dependent acoustic modeling using graphemes for large vocabulary speech recognition

In this paper we propose to use a decision tree based on graphemic acoustic sub-word units together with phonetic questions. We also show that automatic question generation can be used to completely eliminate any manual effort. We present experimental results on four corpora with different languages, namely the Dutch ARISE corpus, the Italian EUTRANS EVAL00 evaluation corpus, the German VERBMOBIL '00 development corpus and the English North American Business '94 20k and 64k development corpora. For all experiments, the acoustic models are trained from scratch in order not to use any prior phonetic knowledge. Complete training procedures have been iterated to simulate the long optimization history used for the phonemic acoustic models. With minimal manual effort we show that for the Dutch, German and Italian corpora, the presented approach works surprisingly well and increases the word error rate by not more than 2% relative. On the English NAB task the error rate is about 20% higher compared to experiments using a pronunciation lexicon.