PASQA: Pitch-Accent-Focused Speech Quality Assessment Model Trained on Synthetic Speech with Accent Errors

2026-06-18 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Speech Processing · Depth: Expert, quick

Summary

PASQA, a novel Pitch-Accent-focused Speech Quality Assessment model, addresses the insensitivity of traditional mean opinion score (MOS) prediction models to localized pitch-accent errors in synthetic speech. Developed to explicitly target pitch-accent correctness, PASQA was trained using a controlled Japanese accent-error dataset. This dataset was generated by modifying accent patterns via an accent-controllable text-to-speech system, with a pseudo accent-quality score derived from the accent-error rate. The model integrates self-supervised representations, mora-conditioned fusion, ranking loss, an auxiliary accent-error localization task, and speaker-invariant training. Experimental results demonstrate PASQA's superior performance, achieving high ordering accuracy on both seen and unseen speakers, a task where conventional models fail to preserve accent-error severity ordering. Furthermore, PASQA exhibits stronger agreement with human accent-correctness judgments. The model's code is publicly available.

Key takeaway

For NLP Engineers or AI Scientists developing text-to-speech systems, especially for pitch-accent languages like Japanese, you should integrate PASQA into your quality assessment pipeline. This model provides superior, fine-grained evaluation of pitch-accent correctness, which traditional MOS models overlook. By adopting PASQA, you can ensure higher fidelity and naturalness in your synthetic speech output, directly addressing a critical aspect of perceived speech quality.

Key insights

PASQA explicitly targets pitch-accent correctness, outperforming general MOS models in localized error detection.

Principles

• Utterance-level MOS models often miss localized pitch-accent errors.
• Training on controlled synthetic accent errors improves accent quality assessment.
• Speaker-invariant training enhances model generalization.

Method

PASQA trains on a Japanese accent-error dataset, generated by an accent-controllable TTS system, using pseudo accent-quality scores. It employs self-supervised representations, mora-conditioned fusion, ranking loss, and an auxiliary accent-error localization task.

In practice

• Use PASQA for fine-grained pitch-accent quality assessment in TTS.
• Leverage accent-controllable TTS for synthetic error dataset generation.
• Integrate speaker-invariant training for robust speech quality models.

Topics

• Speech Quality Assessment
• Pitch Accent
• Text-to-Speech
• Accent Errors
• Self-supervised Learning
• Japanese Language

Code references

• lycorp-jp/PASQA

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Investigating Human-Model Discrepancies in Speech Quality Assessment via Acoustic and Prosodic Perturbations — Current MOS prediction models fail to capture human perception of prosodic errors and speaker characteristics.
• A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition — Fine-tuning pretrained Transformers with specific speech representations and label formats significantly improves Quranic ASR accuracy and efficiency.
• Improving quality and robustness in LLM-based text-to-speech systems — LLM-based TTS quality improves through targeted techniques addressing accent leakage, expressiveness, and reliability.
• KVoiceBench, KOpenAudioBench, and KMMAU: Agent-Driven Korean Speech Benchmarks for Evaluating SpeechLMs — Multilingual SpeechLM evaluation requires language-specific benchmarks to reveal true performance gaps and task-specific weaknesses.
• Perceptual compensation for tonal context in self-supervised speech models — Self-supervised speech models like wav2vec2.0 may require supervised fine-tuning to achieve perceptual compensation for complex phonological features like Mandarin tones.
• ROMPAR: Morphological Completion and Demographic Unlearning for Romanian-Accented Speech Recognition — A multi-task adversarial framework with LLM-guided decoding improves ASR for accented speech by reducing demographic bias and completing morphology.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.