GE2PE: Persian End-to-End Grapheme-to-Phoneme Conversion

Elnaz Rahmati, Hossein SametiEMNLP 2024 · DOI

Introduction

Grapheme-to-Phoneme (G2P) conversion is a critical component of Text-to-Speech (TTS) systems, translating written text into its phonetic representation. For the Persian language, this task is exceptionally challenging due to several linguistic factors: orthographic opacity (where short vowels are omitted), the unwritten Kasre-Ezafe grammatical connector, and the prevalence of homographs that require context for disambiguation. Traditional word-level models often fail to capture the necessary syntactic and semantic relationships to resolve these ambiguities. GE2PE addresses these issues by proposing a sentence-level, end-to-end (E2E) model based on the ByT5 architecture.

Methodology

The GE2PE model utilizes a byte-level transformer (ByT5), which is particularly robust for low-resource languages as it avoids fixed vocabularies and handles out-of-vocabulary words or typos more effectively. To overcome the scarcity of high-quality manually labeled Persian phonetic data, the researchers employed a two-stage "silver-to-gold" training strategy:

  1. Pre-training (Silver Labels): The model is first trained on a massive dataset of 5.3 million sentences with phonemes generated by an existing, albeit imperfect, G2P tool.
  2. Fine-tuning (Gold Labels): The model is then refined using a small, high-precision dataset of 909 manually verified sentences.

This sequential approach allows the model to learn general linguistic structures from large-scale noisy data before specializing on high-quality human annotations, maximizing the impact of the limited manual data.

Data Resources

The study introduces several new resources to the Persian NLP community, ranging from massive machine-generated corpora to specialized evaluation sets for linguistic edge cases.

Table 1: Statistics of the datasets used for training and evaluation, highlighting the scale of the machine-generated silver data compared to the manual gold data.

As shown in the dataset statistics, the "silver" machine-generated data provides the volume necessary for deep learning, while the "gold" FarsDat Aligned set provides the precision. Additionally, the researchers developed specialized test sets—kasre eval and homograph eval—to specifically measure performance on the most difficult aspects of Persian pronunciation.

Experimental Results

GE2PE demonstrates significant improvements over previous state-of-the-art (SOTA) systems. The model achieved a notable reduction in error rates across standard benchmarks.

Table 2: Performance comparison on the FarsDat dataset, showing GE2PE achieving a new state-of-the-art in Word Error Rate (WER) and Phoneme Error Rate (PER).

Key performance highlights include:

  • Word Error Rate (WER): A reduction of 1.86% compared to the previous SOTA (Rezaei et al., 2022).
  • Kasre-Ezafe Detection: A 4.03% improvement in recall, indicating superior handling of unwritten grammatical connectors.
  • Homograph Disambiguation: A 3.42% increase in accuracy for words with context-dependent pronunciations.
  • Efficiency: GE2PE achieves these results with only one-sixth the parameters of previous multi-module SOTA systems, making it more efficient for deployment.

Conclusion

GE2PE represents a major advancement for Persian G2P conversion by successfully applying an end-to-end sentence-level approach. By combining massive silver-label pre-training with targeted gold-label fine-tuning, the model effectively addresses long-standing challenges like Kasre-Ezafe and homographs. While the model currently focuses on the Tehrani accent, future work will aim to expand coverage to regional Iranian accents and explore multi-task learning to further refine its implicit linguistic understanding.

Summary updated

BibTeX
@inproceedings{rahmati2024ge,
  author = {Elnaz Rahmati and Hossein Sameti},
  title = {{GE2PE: Persian End-to-End Grapheme-to-Phoneme Conversion}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.findings-emnlp.196},
}
Figure