Conditional and joint models for grapheme-to-phoneme conversion

By Stanley F. Chen. 2003

186 citations recorded by the source corpus.

In this work, we introduce several models for grapheme-to-phoneme conversion: a conditional maximum entropy model, a joint maximum entropy n -gram model, and a joint maximum entropy n -gram model with syllabification. We examine the relative merits of conditional and joint models for this task, and find that joint models have many advantages. We show that the performance of our best model, the joint n -gram model, compares favorably with the best results for English grapheme-to-phoneme conversion reported in the literature, sometimes by a wide margin. In the latter part of this paper, we consider the task of merging pronunciation lexicons expressed in different phone sets. We show that models for grapheme-to-phoneme conversion can be adapted effectively to this task.

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Tao Qin
2. Xu Tan
3. Tie-Yan Liu
4. Jinyu Li
5. Dong Yu
6. Frank K. Soong
7. Sheng Zhao
8. James R. Glass

Citations in the corpus

• A Survey on Neural Speech Synthesis (2021, methodology): “Thus, the G2P conversion for English is mainly responsible to generate the pronunciations of out-of-vocabulary words [41, 24, 403, 283, 33, 320].”
• Sequence-to-sequence neural net models for grapheme-to-phoneme conversion (2015, methodology): “Our experiments were conducted on the three US English datasets 1 : the CMUDict, NetTalk, and Pronlex datasets that have been evaluated in [18,19].”
• Token-Level Ensemble Distillation for Grapheme-to-Phoneme Conversion (2019, methodology): “Joint sequence n-gram models have been widely used [4, 14, 15] for G2P conversion.”
• Joint-sequence models for grapheme-to-phoneme conversion (2008, methodology): “Chen (2003) uses beam search, whereas Galescu and Allen (2002) do not allow null grapheme units and use simple dynamic programming search.”
• Grapheme-to-phoneme conversion using Long Short-Term Memory recurrent neural networks (2015, background): “We use the same training and testing split as found in [12, 7, 4] and thus the results are directly comparable.”
• Convolutional Sequence to Sequence Model with Non-Sequential Greedy Decoding for Grapheme to Phoneme Conversion (2018, methodology): “Second, we also constructed datasets with stress markings on phonemes, making the task of our study more realistic and difficult than those of previous studies, which removed the stress markings [5, 14, 25].”
• A bi-directional LSTM approach for polyphone disambiguation in Mandarin Chinese (2016, background): “In the G2P conversion task [1, 2, 3, 4, 5, 6, 7], a neural network accepts a sequence of characters and outputs a sequence of pronunciations.”
• Text-to-Speech Synthesis (2009, methodology): “There areM2 di-syllables.[86], [277] syllables Syllables as normally defined [236] [388] [512] words Words as normally defined [480], [416], [359].”
• Encoding linear models as weighted finite-state transducers (2014, unknown): “This is the same evaluation setup as found in [20, 21, 10, 11].”
• Learning Lexicons From Speech Using a Pronunciation Mixture Model (2013, methodology): “For example, the phrase “nearest tornado” might be represented at the phoneme level as (1) in the table below, but with the application of phonological rules the phone pronunciation might read as either (2) or (3).”