Phonetic pronunciations for arabic speech-to-text systems

Abstract

In this paper two aspects of generating and using phonetic Arabic dictionaries are described. First, the use of single pronunciation acoustic models in the context of Arabic large vocabulary Automatic Speech Recognition (ASR) is investigated. These have been found to be useful for English ASR systems, when combined with standard multiple pronunciation systems. The second area examined is automatically deriving phonetic "pronunciations" for words that standard approaches, such as the Buckwalter Morphological Analyzer, cannot handle. Without pronunciations for these words the OOV rates for various Arabic tasks significantly increase. Here, pronunciations are automatically found by first deriving grapheme-to-phone rules, and associated rule probabilities. These are then used to produce the most likely pronunciation, or pronunciations, for any word. These approaches are evaluated on a large vocabulary Arabic Broadcast News and Broadcast Conversation transcription task. Both schemes are found to yield gains with a multi-pass/combination framework.

Comparison of unsupervised classification methods for brain tumor segmentation using multi-parametric MRI

“…Hard segmentation is obtained by assigning each voxel to the tissue class with the highest probability. In order to recover from singular covariance matrices, a regularization constant of 1e − 3 was added to the diagonal elements of the covariance matrices (Diehl et al, 2011). …”

Efficient generation and use of MLP features for Arabic speech recognition

“…Pass three performs acoustic lattice rescoring, applying constrained Maximum Likelihood Linear Regression (MLLR) followed by lattice-MLLR and Confusion Network (CN) decoding. Further details on the training and decoding configuration can be found in [9].…”

Training and adapting MLP features for Arabic speech recognition

“…Out-of-vocabulary rates for the 260k and 350k wordlists is the subset of the 350K word-list for which phonetic pronunciations could be obtained using Buckwalter. The 90K missing pronunciations were found using automatically derived rules described in [11]. The out-of-vocabulary (OOV) rates for the three test sets used are shown in table 2.…”

“…The second, a phonetic system (V1), was based on 39 phones, the graphemic ones plus the three short vowels. For further details of the two systems see [11]. Both models used a 39-dimensional PLP-based front-end which used 13 PLP cepstra, including the zeroth cepstral coefficient with first, second and third delta parameters appended followed by an HLDA projection from 52-dimensions down to 39.…”

“…For the phonetic training pronunciations Buckwalter was used. Any missing pronunciations from the acoustic model training data were obtained using the automatic pronunciation system described in [11].…”