Posts by Collection

Portfolio item number 1

Short description of portfolio item number 1

Portfolio item number 2

Short description of portfolio item number 2

Multi-Modal Automatic Prosody Annotation with Contrastive Pretraining of Speech-Silence and Word-Punctuation

Jinzuomu Zhong, Yang Li, Hui Huang, Korin Richmond, Jie Liu, Zhiba Su, Jing Guo, Benlai Tang, Fengjie Zhu

Published in Interspeech 2024

🛠️ Conducted research and engineering on multimodal prosody annotation during my Transsion internship, applying the approach to an existing data curation pipeline.

💡 A multimodal annotator contrastively pretrained on speech-silence and word-punctuation pairs achieves absolute prosodic-word F1 gains of 0.20–0.29 over a multimodal baseline. It remains effective with limited labelled data and improves downstream TTS listener preference.

Paper | arXiv | Demo | Code

AccentBox: Towards High-Fidelity Zero-Shot Accent Generation

Jinzuomu Zhong, Korin Richmond, Zhiba Su, Siqi Sun

Published in ICASSP 2025

🛠️ Developed an embedding-based approach to accent conditioning in my MSc thesis, connecting speaker-generalising accent identification with zero-shot TTS to address accent hallucination.

💡 GenAID reduces speaker-accent entanglement to generalise accent identification to unseen speakers; its continuous embeddings improve accent fidelity in zero-shot TTS. In cross-accent generation, listeners prefer AccentBox over discrete accent-label conditioning in 61.7–70.0% of comparisons across Irish and American accents.

Paper | arXiv | Demo | Code | MSc Thesis

Pairwise Evaluation of Accent Similarity in Speech Synthesis

Jinzuomu Zhong, Suyuan Liu, Dan Wells, Korin Richmond

Published in Interspeech 2025

🛠️ Led a collaboration in phonetics and TTS on accent-similarity evaluation, combining listening-test refinements with pronunciation-based objective metrics.

💡 Transcript-guided XAB tests with accent-difference highlighting and listener screening detect system differences with as few as five valid listeners in an Edinburgh-accent case study. Formant and phonetic-posteriorgram distances track hypothesised system degradation more consistently than WER or predicted naturalness.

Paper | arXiv

Rethinking Discrete Speech Representation Tokens for Accent Generation

Jinzuomu Zhong*, Yi Wang*, Korin Richmond, Peter Bell

Preprint

🛠️ Co-led a study of discrete speech representation tokens for controllable generation, comparing encoder layers, ASR supervision, and quantisation through Accent ABX and cross-accent voice conversion.

💡 Accent recoverability peaks in mid-early HuBERT layers (6 and 9), rather than the deeper layers commonly used for speech tokens. Accent ABX and cross-accent resynthesis show that ASR supervision suppresses accent cues, while codebook reduction alone does not selectively disentangle them.

arXiv | Demo

Accent-Emotion Entanglement in LM-Based Text-to-Speech Systems

Matthew Hayden, Jinzuomu Zhong, Korin Richmond

Published in Interspeech 2026

🛠️ Co-supervised an MSc research project investigating accent-emotion entanglement in zero-shot TTS through targeted perceptual evaluation and accent-representation analysis.

💡 Emotion conditioning can shift accent in CosyVoice 2 despite strong reported speaker-similarity scores, while MaskGCT is more stable in the same case study. Accent-focused listening tests and embedding analysis reveal a controllability failure that conventional speaker metrics can overlook.

Paper | Demo

CycleCodec: Distillation-Free Factorized Neural Speech Codec via Cycle-Consistent Speaker Swapping

Rui-Chen Zheng, Nicholas Sanders, Jinzuomu Zhong, Yang Ai, Zhen-Hua Ling, Korin Richmond

Published in Interspeech 2026

🛠️ Contributed to a visiting PhD student’s project on factorised acoustic tokens, investigating speaker-content disentanglement and cross-lingual voice conversion without pretrained teacher supervision.

💡 Cycle-consistent speaker swapping improves content preservation and speaker transfer without distillation from pretrained SSL or ASR teachers. Trained on English, CycleCodec outperforms TiCodec on English, Mandarin, and Vietnamese voice conversion and accumulates less content drift under repeated conversion.

Paper

When Text-to-Speech Speaks in Your Voice: A Study on Public Perception

Ariadna Sanchez*, Jinzuomu Zhong*, Artemis Deligianni*, Alice Ross, Simon King

Published in ACM CUI 2026

🛠️ Co-led public engagement activities and an own-voice perception study, bringing voice cloning into live demonstrations to investigate how people assess synthetic versions of their speech.

💡 In a 72-participant study, listeners distinguish recordings of their own voices from MaskGCT clones with an estimated 87% accuracy, citing pitch, accent, and intonation. The findings motivate evaluating personalised TTS with voice owners alongside conventional ratings of unfamiliar voices.

Paper

Beyond Utterance-Level Speaker Similarity: Perception-Aligned Segmental Evaluation of Accent in Speech Generation

Jinzuomu Zhong, Alice Ross, Catherine Lai, Korin Richmond

Preprint

🛠️ Led a collaboration in phonetics and sociolinguistics on segmental accent evaluation, producing an open evaluation dataset, analysis code, and an interactive results demo.

💡 Subword annotations from target-accent listeners reveal pronunciation differences and system rankings that holistic similarity ratings can obscure. Across Scottish and Australian voices, seven objective distances correlate only weakly with local judgements (Spearman ρ ≤ 0.22), exposing a gap in current speech-generation evaluation.

arXiv | Demo | Code

Accent Generation - Challenges and Progress

Published: April 11, 2025

“AI for All” is a special public engagement event with interactive exhibitions, live demonstrations, and student-led activities. This event is hosted by the Centre for Doctoral Training in Designing Responsible Natural Language Processing together with Edinburgh Science Festival and Edinburgh Futures Institute.

Checkout my Demo and Poster (thanks to the brilliant design by Prof. John Vines)!

Accent Hallucination in Zero-Shot Text-to-Speech

Published: June 15, 2025

UKRI Centres for Doctoral Training in NLP Partners Summit 2025 is a cross-sectoral event for doctoral students, leading academics, and industry partners, showcasing cutting-edge NLP research with research posters and project presentations. This event is hosted by the Centre for Doctoral Training in Designing Responsible Natural Language Processing and Centre for Doctoral Training in Natural Language Processing at the University of Edinburgh.

Checkout my Poster!

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.