Improving Proper Name Recognition by Adding Automatically Learned Pronunciation Variants to the Lexicon

This paper deals with the task of large vocabulary proper name recognition. In order to accomodate a wide diversity of possible name pronunciations (due to non-native name origins or speaker tongues) a multilingual acoustic model is combined with a lexicon comprising 3 grapheme-to-phoneme (G2P) transcriptions (from G2P transcribers for 3 different languages) and up to 4 so-called phoneme-to-phoneme (P2P) transcriptions. The latter are generated with (speaker tongue, name source) specific P2P converters that try to transform a set of baseline name transcriptions into a pool of transcription variants that lie closer to the 'true' name pronunciations. The experimental results show that the generated P2P variants can be employed to improve name recognition, and that the obtained accuracy is comparable to what is achieved with typical (TY) transcriptions (made by a human expert). Furthermore, it is demonstrated that the P2P conversion can best be instantiated from a baseline transcription in the name source language, and that knowledge of the speaker tongue is an important input as well for the P2P transcription process.