Elamur Research⌕
☰
home / papers / 2026 / Beyond Unified Models: A Service-Oriented Approach to Low Latency, Context Aware Phonemization for Real Time TTS

Beyond Unified Models: A Service-Oriented Approach to Low Latency, Context Aware Phonemization for Real Time TTS

Paper · 2026 · Mahta Fetrat Qharabagh, Donya Navabi, Zahra Dehghanian et al.

Summary

Fast on-device TTS engines such as Piper separate grapheme-to-phoneme (G2P) conversion from neural phoneme-to-speech synthesis. Their lightweight rule-based phonemizers (e.g., eSpeak-ng) cannot resolve context-dependent pronunciation, which is a serious problem for Persian, where homographs and the unwritten Ezafe vowel /e/ can change meaning. Stronger neural G2P models fix this but are too slow for real-time, CPU-only use cases such as screen readers.

The authors address this in two ways. First, they build cheaper context-aware components, including an enhanced eSpeak-based Persian phonemizer and an ALBERT-based Ezafe detector distilled from a SpaCy teacher. Second, they restructure Piper so that these modules run as separate services that the core engine queries, rather than loading them inside the synthesis runtime. They also release a new Persian Piper voice trained on what they describe as the largest public Persian TTS dataset.

On a CPU-only laptop, the service-based setup keeps real-time speed while greatly improving Ezafe and homograph handling and listener ratings. Code and models are open-source.

Key points

  • Accuracy gains: Compared with baseline Piper, Piper + LCA G2P lowers PER from 6.32% to 4.80%. It raises Ezafe F1 from 19.58% to 90.08% and homograph accuracy from 43.87% to 77.67%. Piper + Neural G2P, which uses the 300M-parameter Homo-GE2PE, reaches 4.95% PER, 87.70% Ezafe F1 and 74.53% homograph accuracy.
  • Speed (RTF, Intel i7-1255U, CPU only): Direct integration is slower than real time, with an RTF of 5.519 for LCA and 3.840 for Neural G2P. Running the modules as services brings RTF down to 0.167 and 0.396 respectively, close to base Piper's 0.153.
  • Listener ratings: In a MOS test with 16 native Persian speakers rating seven utterances, Piper + LCA G2P scored 3.14. This beats base Piper (2.41), MatchaTTS (2.54) and GlowTTS (1.30); natural speech scored 4.21.
  • Distilled Ezafe detector: The ALBERT-based model has 11.09M parameters and uses 42.32 MB of memory, compared with 162.84M parameters and 621.19 MB for the SpaCy teacher. It keeps an Ezafe F1 of 94.19% (teacher: 97.67%) and runs about three times faster (0.037 s vs 0.110 s).
  • Release: Code, models and results are published at github.com/MahtaFetrat/Piper-with-LCA-Phonemizer.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Sharif University of TechnologyTehran · Iran5 works · 524 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Mahta Fetrat Qharabagh, Donya Navabi, Zahra Dehghanian, Morteza Abolghasemi and Hamid R. Rabiee. 2026. Beyond Unified Models: A Service-Oriented Approach to Low Latency, Context Aware Phonemization for Real Time TTS. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-industry.12

@inproceedings{qharabagh2026beyond,
  title = {{Beyond Unified Models: A Service-Oriented Approach to Low Latency, Context Aware Phonemization for Real Time TTS}},
  author = {Mahta Fetrat Qharabagh, Donya Navabi, Zahra Dehghanian, Morteza Abolghasemi and Hamid R. Rabiee},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-industry.12},
  url = {https://doi.org/10.18653/v1/2026.eacl-industry.12}
}

References

51 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mohammad Hasan Sohan Ajini. 2022. Attention based grapheme to phoneme. Accessed: 2025-11-17. find
  2. Sajad Alipour. 2023. Persian grapheme to phoneme with transformer. Accessed: 2025-11-17. find
  3. Kenneth R Beesley and Lauri Karttunen. 2003. Finite-state morphology: Xerox tools and techniques. CSLI, Stanford, pages 359–375. find
  4. Maximilian Bisani and Hermann Ney. 2008. Joint-sequence models for grapheme-to-phoneme conversion. Speech communication, 50(5):434–451. find
  5. Alan W. Black, Paul Taylor, and 1 others. 2004. Festival Speech Synthesis System: Server/Client API. Section 28.3: BSD socket server; long-lived synthesis process. find
  6. Alan W Black and Paul A Taylor. 1997. Automatically clustering similar units for unit selection in speech synthesis. find
  7. Hafez Dehghani. 2022. persian_phonemizer: A tool for translating persian text to ipa. Accessed: 2025-11-17. find
  8. Homer Dudley, Richard R Riesz, and Stanley SA Watkins. 1939. A synthetic speaker. Journal of the Franklin Institute, 227(6):739–764. eSpeak NG developers. 2013. espeak NG text-to-speech engine. GitHub repository. Accessed: 2025-11-17. find
  9. Mahta Fetrat. 2025a. Homo-ge2pe-persian: Persian grapheme-to-phoneme conversion with homograph disambiguation. Accessed: 2025-11-08. find
  10. Mahta Fetrat. 2025b. Homofast-espeak-persian: A homograph-aware persian g2p extension of espeak ng. Accessed: 2025-11-08. find
  11. HooshvareLab. 2021. Albert-persian: A lite bert for self-supervised learning of language representations for the persian language. Model: albert-fa-zwnj-base-v2. Accessed: 2025-11-07. find
  12. Andrew J Hunt and Alan W Black. 1996. Unit selection in a concatenative speech synthesis system using a large speech database. In 1996 IEEE international conference on acoustics, speech, and signal processing conference proceedings, volume 1, pages 373–376. IEEE. find
  13. Sittichai Jiampojamarn, Grzegorz Kondrak, and Tarek Sherif. 2007. Applying many-to-many alignments and hidden markov models to letter-to-phoneme conversion. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings…
  14. Kamtera. 2023. persian-tts-female-glow_tts: Single-speaker female persian text-to-speech model. Accessed: 2025-11-07. find
  15. Sadegh Karimi. 2024. persian-text-to-speech: Persian text-to-speech model. Accessed: 2025-11-17. find
  16. Heeseung Kim, Sungwon Kim, and Sungroh Yoon. 2022. Guided-tts: A diffusion model for text-to-speech via classifier guidance. In International Conference on Machine Learning, pages 11119–11133. PMLR. find
  17. Jaehyeon Kim, Sungwon Kim, Jungil Kong, and Sungroh Yoon. 2020. Glow-tts: A generative flow for text-to-speech via monotonic alignment search. Advances in Neural Information Processing Systems, 33:8067–8077. find
  18. Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In International Conference on Machine Learning, pages 5530–5540. PMLR. find
  19. Dennis H Klatt. 1980. Software for a cascade/parallel formant synthesizer. the Journal of the Acoustical Society of America, 67(3):971–995. find
  20. Dennis H Klatt. 1987. Review of text-to-speech conversion for english. The Journal of the Acoustical Society of America, 82(3):737–793. find
  21. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942.
  22. Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, and 1 others. 2023. Voicebox: Text-guided multilingual universal speech generation at scale. Advances in neural information processing systems, 36:14005–14034. find
  23. Yinghao Aaron Li, Cong Han, and Nima Mesgarani. 2025. Styletts: A style-based generative model for natural and diverse text-to-speech synthesis. IEEE Journal of Selected Topics in Signal Processing. find
  24. Shivam Mehta, Ruibo Tu, Jonas Beskow, Éva Székely, and Gustav Eje Henter. 2024. Matcha-tts: A fast tts architecture with conditional flow matching. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11341–11345. IEEE. find
  25. David R Mortensen, Siddharth Dalmia, and Patrick Littell. 2018. Epitran: Precision g2p for many languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).
  26. Amir Hossein Navabi. 2025. persian-tts-piper: A hugging face space for persian text-to-speech using piper. Accessed: 2025-11-07. find
  27. OHF-Voice. 2025. piper1-gpl: Fast and local neural text-to-speech engine. Accessed: 2025-11-07. find
  28. Musharraf Omer. 2025. sonata-nvda: A speech synthesizer driver for nvda using neural tts models. Accessed: 2025-11-07. find
  29. Kyubyong Park. 2019. g2pE: A simple english grapheme-to-phoneme converter (seq2seq/transformer implementation). GitHub repository. find
  30. Demetry Pascal. 2020. Simple persian (farsi) grapheme-to-phoneme converter. Accessed: 2025-11-17. find
  31. Pipecat AI. 2024. Piperttsservice: Self-hosted httpserver for piper tts. Product documentation. Accessed: 2025-11-17. find
  32. Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. 2021. Grad-tts: A diffusion probabilistic model for text-to-speech. In International conference on machine learning, pages 8599–8608. PMLR. find
  33. Ryan Prenger, Rafael Valle, and Bryan Catanzaro. 2019. Waveglow: A flow-based generative network for speech synthesis. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3617–3621. IEEE. find
  34. Mahta Fetrat Qharabagh, Zahra Dehghanian, and Hamid R Rabiee. 2025a. Fast, not fancy: Rethinking g2p with rich data and rule-based models. arXiv preprint arXiv:2505.12973. — [arXiv:2505.12973](https://arxiv.org/abs/2505.12973) find
  35. Yinghao Aaron Li, Cong Han, Vinay Raghavan, Gavin Mischler, and Nima Mesgarani. 2023. Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models. Advances in Neural Information Processing Systems, 36:19594–19621. find
  36. LlamaEdge contributors. 2024. Tts-api-server: Restful api server for piper (openai-compatible routes). GitHub repository. Accessed: 2025-11-17. find
  37. Ali Mahmoudi. 2025. Khadijah-fa_en-matcha-tts-model: A persian/english text-to-speech model using matcha-tts. Accessed: 2025-11-17. find
  38. MARYTTS. 2022. Mary tts: an open-source, multilingual text-to-speech synthesis system. Accessed: 2025-11-16. find
  39. Mahta Fetrat Qharabagh, Zahra Dehghanian, and Hamid R Rabiee. 2025b. Llm-powered grapheme-to-phoneme conversion: Benchmark and case study. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. find
  40. Mahta Fetrat Qharabagh, Zahra Dehghanian, and Hamid R Rabiee. 2025c. Manatts persian: a recipe for creating tts datasets for lower resource languages. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technol… find
  41. Azam Rabiee. 2019. Persian g2p. GitHub repository. Accessed: 2025-11-17. find
  42. Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2020. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv preprint arXiv:2006.04558. find
  43. Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. 2020. Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6199–6203.… find
  44. Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2019. Fastspeech: Fast, robust and controllable text to speech. Advances in neural information processing systems, 32. find
  45. Kaisheng Yao and Geoffrey Zweig. 2015. Sequence-to-sequence neural net models for grapheme-to-phoneme conversion. arXiv preprint arXiv:1506.00196. — [arXiv:1506.00196](https://arxiv.org/abs/1506.00196) find
  46. Rhasspy Team. 2023. Piper: Fast, local neural text to speech. GitHub repository. find
  47. ZachB100. 2023. Piper training guide with screen reader. GitHub repository. Accessed: 2025-11-17. find
  48. Roshan. 2023. spacy_pos_tagger_parsbertpostagger. Accessed: 2025-11-13. find
  49. Heiga Zen, Keiichi Tokuda, and Alan W Black. 2009. Statistical parametric speech synthesis. speech communication, 51(11):1039–1064. find
  50. Yoshinori Sagisaka. 1988. Speech synthesis by rule using an optimal selection of non-uniform synthesis units. In ICASSP-88., International Conference on Acoustics, Speech, and Signal Processing, pages 679– 682. IEEE. find
  51. A Jonathan Shen, Ruoming Pang, Ron J Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, Rj Skerrv-Ryan, and 1 others. 2018. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions. In 2018 IEEE international conference on acoustics, speech… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/qharabagh-2026-beyond-unified-models-service-oriented/