Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Mahta Fetrat, Donya Navabi, Zahra Dehghanian, Morteza Abolghasemi, Hamid R. Rabiee

TL;DR

The paper tackles the latency-accuracy tension in G2P-aided real-time TTS by introducing a service-oriented framework that decouples heavy context-aware phonemization from the core TTS engine. It combines lightweight statistical context methods with a distilled Ezafe detector and an IPC-based architecture to enable heavier neural G2P components without compromising real-time performance. Key contributions include a Persian G2P front-end enhanced for homographs and Ezafe, a PiperTTS-based workflow with a Persian voice, and CPU-only real-time demonstrations on end devices. The results show improved pronunciation and linguistic accuracy while preserving real-time responsiveness, aided by open-source release of resources.

Abstract

Lightweight, real-time text-to-speech systems are crucial for accessibility. However, the most efficient TTS models often rely on lightweight phonemizers that struggle with context-dependent challenges. In contrast, more advanced phonemizers with a deeper linguistic understanding typically incur high computational costs, which prevents real-time performance. This paper examines the trade-off between phonemization quality and inference speed in G2P-aided TTS systems, introducing a practical framework to bridge this gap. We propose lightweight strategies for context-aware phonemization and a service-oriented TTS architecture that executes these modules as independent services. This design decouples heavy context-aware components from the core TTS engine, effectively breaking the latency barrier and enabling real-time use of high-quality phonemization models. Experimental results confirm that the proposed system improves pronunciation soundness and linguistic accuracy while maintaining real-time responsiveness, making it well-suited for offline and end-device TTS applications.

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

TL;DR

The paper tackles the latency-accuracy tension in G2P-aided real-time TTS by introducing a service-oriented framework that decouples heavy context-aware phonemization from the core TTS engine. It combines lightweight statistical context methods with a distilled Ezafe detector and an IPC-based architecture to enable heavier neural G2P components without compromising real-time performance. Key contributions include a Persian G2P front-end enhanced for homographs and Ezafe, a PiperTTS-based workflow with a Persian voice, and CPU-only real-time demonstrations on end devices. The results show improved pronunciation and linguistic accuracy while preserving real-time responsiveness, aided by open-source release of resources.

Abstract

Lightweight, real-time text-to-speech systems are crucial for accessibility. However, the most efficient TTS models often rely on lightweight phonemizers that struggle with context-dependent challenges. In contrast, more advanced phonemizers with a deeper linguistic understanding typically incur high computational costs, which prevents real-time performance. This paper examines the trade-off between phonemization quality and inference speed in G2P-aided TTS systems, introducing a practical framework to bridge this gap. We propose lightweight strategies for context-aware phonemization and a service-oriented TTS architecture that executes these modules as independent services. This design decouples heavy context-aware components from the core TTS engine, effectively breaking the latency barrier and enabling real-time use of high-quality phonemization models. Experimental results confirm that the proposed system improves pronunciation soundness and linguistic accuracy while maintaining real-time responsiveness, making it well-suited for offline and end-device TTS applications.

Paper Structure

This paper contains 15 sections, 1 equation, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. TTS approaches
  4. G2P Tools
  5. Decoupled TTSs
  6. Methodology
  7. Statistical Context-Awareness
  8. Distilled Linguistic Knowledge
  9. Service-Based Integration
  10. Experiments
  11. Mean Opinion Score
  12. Ezafe Detection Module Evaluation
  13. Conclusion
  14. Selection Criteria for PiperTTS
  15. Mean Opinion Score Details

Figures (6)

  • Figure 1: An example of how the Persian Ezafe phoneme (/e/) can change the meaning of a sentence.
  • Figure 2: Four common granularity levels in TTS architectures, differing by intermediate representations.
  • Figure 3: The proposed service-based architecture for context-aware TTS.
  • Figure 4: Overall average MOS across all seven utterances for each TTS system and natural speech, highlighting the improved naturalness of the enhanced Piper system.
  • Figure 5: Distribution of MOS ratings assigned by participants to the enhanced TTS system (Piper + LCA-G2P).
  • ...and 1 more figures