Models for ASR
- Models are trained on the si-84 subset of the LDC WSJ0 corpus. Both MFCC and PLP models are provided, with two phonetic representations (gam and arpabet) based on the Unilex dictionary. In addition, the MFCC arpabet models are provided with the AMR codec. The release is subject to the LDC conditions; that is, the acoustic models are released as a derived work, but for research only.
- The Mandarin FSTs are created by Nokia. The correpsonding acoustic models are trained with Mandarin PCOM database. The language model is a direct translation of the Finnish Phrasebook language model. The FSTs can be used for research only.
- Japanese News paper Article Sentences (JNAS) database were used for training HMMs. To train an acoustic model for speech recognition was trained from 37k sentences uttered by 122 male and 122 female speakers. And the language models used for speech recognition were based on Mainichi news paper courpus.