arXiv · 2025 · Preprint
Fetrat et al. (Sharif University of Technology) · → Paper · Demo: ✓ · Code: ✓
Proposes a service-oriented TTS architecture that runs heavy context-aware phonemization as an independent process alongside a lightweight real-time synthesizer, restoring low latency while substantially improving Persian pronunciation accuracy over the underlying rule-based phonemizer.
Problem
Lightweight, real-time TTS systems intended for offline, on-device use (e.g., screen readers) must satisfy three competing requirements: lightweightness, real-time responsiveness, and naturalness. A common strategy for meeting the first two is decoupling grapheme-to-phoneme (G2P) conversion from phoneme-to-speech synthesis and using a fast rule-based phonemizer. But rule-based phonemizers fail on context-dependent phenomena, and for Persian specifically this includes homograph disambiguation (words with multiple valid pronunciations depending on context) and Ezafe detection (an unwritten connecting /e/ sound between grammatically related words whose presence can change a sentence’s meaning). Stronger, context-aware neural phonemizers exist but are too large and slow to embed directly into a real-time TTS runtime without breaking the latency budget the deployment scenario requires.
Method
The baseline system is Piper, a VITS-derived, ONNX-exported TTS engine that separates a rule-based eSpeak G2P front-end from a neural phoneme-to-speech (P2S) back-end. The paper adds two complementary phonemization enhancements and a systems-level integration mechanism to keep them fast. First, a lightweight statistical homograph disambiguator (adapted from prior work by the same author group) selects a pronunciation using word co-occurrence context, without any neural inference. Second, an Ezafe detector is produced by distilling a large Persian SpaCy part-of-speech tagger (162.84M parameters, treating Ezafe detection as a POS-tagging subtask) into a much smaller ALBERT-based model (11.09M parameters), exported to ONNX for CPU inference.
Rather than embedding these context-aware modules directly inside the TTS runtime, where their load and inference costs would accumulate on every request, the paper runs them as an independent, persistent service that the core TTS engine communicates with over piped inter-process communication (IPC). For each input, the core engine first generates an initial phoneme sequence with its default eSpeak phonemizer, sends it to the phonemization service for homograph correction and Ezafe insertion, and receives back an enhanced sequence that is passed to the P2S model. The P2S model itself is then fine-tuned (1,000 epochs, ManaTTS Persian dataset) on phoneme sequences produced by this enhanced phonemizer, so the acoustic model learns to handle the corrected Ezafe phonemes and disambiguated homographs rather than defaulting to majority-class pronunciations.
Key Results
The service-based architecture removes nearly all of the latency the stronger phonemizers would otherwise add: real-time factor (RTF) for the enhanced lightweight system (“Piper + LCA G2P”) drops from 5.519 when called directly to 0.167 when run as a service, and for the heavier 300M-parameter neural G2P variant (“Piper + Neural G2P”) RTF drops from 3.840 to 0.396, both well within real-time budgets. Phonemization accuracy improves substantially over the plain eSpeak-based Piper baseline: Ezafe F1 rises from 19.58% to 90.08% and homograph disambiguation accuracy from 43.87% to 77.67% with the lightweight enhancements, matching or exceeding the much larger neural G2P’s 87.70% Ezafe F1 and 74.53% homograph accuracy. The distilled ALBERT Ezafe tagger retains most of its SpaCy teacher’s accuracy (94.19% vs. 97.67% F1) while cutting parameters by roughly 15x and average inference time from 0.110s to 0.037s. In a 16-rater MOS listening test over 7 held-out utterances, the enhanced system scored 3.14 ± 1.00 versus 2.41 ± 0.84 for baseline Piper, though both remain below the 4.21 ± 0.97 MOS of natural speech recordings.
Novelty Assessment
The individual phonemization components build directly on the authors’ own prior published work (statistical homograph disambiguation and the underlying Homo-GE2PE/HomoFast-eSpeak G2P line from earlier papers); the genuinely new pieces here are the ALBERT-based Ezafe distillation and, more centrally, the service-oriented decomposition of the TTS pipeline itself. The paper explicitly frames this service-based design, treating latency-heavy utility modules as independently-running processes queried over IPC rather than embedded in the synthesis runtime, as the first such structuring of TTS internals (as distinct from prior work exposing an entire synthesizer as a remote API). The contribution is a systems/integration pattern applied to a specific, well-motivated deployment constraint rather than a new generative architecture.
Field Significance
moderate — The paper addresses a genuine and underexplored engineering trade-off (context-aware phonemization quality vs. real-time, on-device latency) with a reusable architectural pattern, service-based decoupling of heavy TTS submodules, validated with both objective (RTF, PER, Ezafe F1, homograph accuracy) and subjective (16-rater MOS) measurements on a concrete deployment target (Persian screen readers). Its scope is narrow (one baseline architecture, one language), but the underlying service-decoupling idea is not specific to Persian or to Piper and could generalize to other lightweight TTS pipelines facing the same latency/quality tension.
Claims
- supports: Decoupling computationally heavy auxiliary TTS components into independently-running services accessed over inter-process communication can remove most of the latency they would otherwise add if embedded directly in the synthesis runtime, without sacrificing their accuracy gains.
Evidence: RTF for the enhanced lightweight phonemizer drops from 5.519 (direct call) to 0.167 (service-based), and for the larger neural G2P variant from 3.840 to 0.396, with phonemization accuracy metrics unchanged between the two calling modes. (§4, Table 1)
- supports: Lightweight, non-neural context-aware enhancements to a rule-based G2P module can close most of the accuracy gap to a much larger neural G2P model on context-dependent phonemization subtasks.
Evidence: A statistical homograph disambiguator and a distilled Ezafe tagger raise Ezafe F1 from 19.58% to 90.08% and homograph accuracy from 43.87% to 77.67%, matching a 300M-parameter neural G2P’s 87.70% Ezafe F1 and 74.53% homograph accuracy. (§4, Table 1)
- supports: Knowledge distillation from a large general-purpose linguistic tagging model into a small task-specific model can retain most of the teacher’s accuracy on a narrow phonemization subtask while cutting parameter count and inference time by roughly an order of magnitude.
Evidence: An ALBERT-based Ezafe tagger distilled from a 162.84M-parameter SpaCy POS tagger uses 11.09M parameters and 0.037s average inference time versus the teacher’s 0.110s, with F1 dropping only from 97.67% to 94.19%. (§3.2, Table 2)
- complicates: Correcting a TTS system’s phoneme input can meaningfully improve perceived naturalness without closing the gap to natural human speech, because naturalness also depends on prosodic and expressive capacity that the phonemizer does not control.
Evidence: Fine-tuning on enhanced phoneme sequences raised MOS from 2.41 to 3.14 over baseline Piper, but this remained well below the 4.21 MOS of natural speech, which the authors attribute to the lightweight P2S model’s limited capacity to reproduce higher-level prosodic and expressive features. (§4.1, §Limitations, Table 3)
Limitations and Open Questions
Even with fully corrected phoneme sequences, the reported naturalness gain (MOS 2.41 to 3.14) leaves a substantial gap to natural speech (MOS 4.21); the authors attribute this to the lightweight P2S model's limited capacity for prosody and expressiveness rather than to remaining phonemization errors, meaning the paper's method addresses pronunciation soundness but not overall naturalness.
The authors note that naturalness is more closely tied to smoothness, noiselessness, and prosody than to phonemization correctness per se, and suggest that subjective evaluation protocols should separate these dimensions rather than conflating them. They also flag headroom in the service layer itself, request-level parallelism or asynchronous processing are not yet implemented and could further reduce latency. The evaluation is limited to a single baseline architecture (Piper/VITS) and a single language (Persian) with a small 16-rater, 7-utterance MOS study.
Wiki Connections
- Subjective Evaluation — validates the naturalness improvement from enhanced phonemization with a 16-rater MOS listening test against baseline Piper and natural speech references.
- Evaluation Metrics — defines a composite phonemization-quality metric (combining Ezafe F1, homograph accuracy, and phoneme error rate) to visualize the speed/quality trade-off across systems.