Chang ZENG 曾畅 曾 暢 (ソウ チョウ)
Independent Researcher
Biography
Independent researcher in generative audio and voice LLMs with 7+ years of experience from research to production. I have led end-to-end development of expressive TTS systems and contributed to full-duplex speech systems across data curation, codec and tokenization design, large-scale multi-GPU training, and deployment. My work focuses on turning advanced speech and audio methods into product-ready AI systems for voice generation, multimodal interaction, and intelligent audio understanding.
Download my resumé .
Interests
- Generative Audio and Voice LLMs
- Multimodal Foundation Models
- Speech and Singing Voice Generation
- Speaker Recognition and Antispoofing
- Audio Separation and Enhancement
Education
PhD in Informatics, 2024
National Institute of Informatics & SOKENDAI
MEng in Electrical Engineering and Information System, 2020
The University of Tokyo
BEng in Measurement and Control Technology and Instruments, 2016
Tianjin University
News
- 2026.06: Speech Codec Probing from Semantic and Phonetic Perspectives accepted by Interspeech 2026.
- 2026.05: Released StepAudio 2.5 Technical Report on arXiv.
- 2026.05: A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation accepted by ICML 2026.
- 2026.03: Released BEAVER: A Training-Free Hierarchical Prompt Compression Method via Structure-Aware Page Selection on arXiv.
- 2026.01: DrivingScene and PAGS accepted by ICASSP 2026.
- 2025.12: Released Towards Interactive Intelligence for Digital Humans on arXiv.
- 2025.10: Critical Information Only accepted by IEEE TDSC.
- 2025.01: SonicSim accepted by ICLR 2025.
- 2025.01: A Benchmark for Multi-Speaker Anonymization accepted by IEEE TIFS.
- 2024.12: InstructSing accepted by SLT 2024.
- 2024.12: Spoofing-Aware Speaker Verification Robust Against Domain and Channel Mismatches accepted by SLT 2024.
- 2024.03: Released HAM-TTS on arXiv.
Publications
Quickly discover relevant content by filtering publications.
(2026).
Speech Codec Probing from Semantic and Phonetic Perspectives.
Accepted by Interspeech 2026.
(2026).
StepAudio 2.5 Technical Report.
In arXiv.
(2026).
A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation.
Accepted by ICML 2026.
(2026).
BEAVER: A Training-Free Hierarchical Prompt Compression Method via Structure-Aware Page Selection.
In arXiv.
(2026).
DrivingScene: A Multi-Task Online Feed-Forward 3D Gaussian Splatting Method for Dynamic Driving Scenes.
Accepted by ICASSP 2026.
(2026).
PAGS: Priority-Adaptive Gaussian Splatting for Dynamic Driving Scenes.
Accepted by ICASSP 2026.
(2025).
Towards Interactive Intelligence for Digital Humans.
In arXiv.
(2025).
Critical Information Only: A Content Privacy-Preserving Framework for Detecting Audio Deepfakes.
In IEEE TDSC.
(2025).
SonicSim: A Customizable Simulation Platform for Speech Processing in Moving Sound Source Scenarios.
Accepted by ICLR 2025.
(2025).
A Benchmark for Multi-Speaker Anonymization.
In IEEE TIFS.
(2024).
InstructSing: High-Fidelity Singing Voice Generation via Instructing Yourself.
In SLT 2024.
(2024).
Spoofing-Aware Speaker Verification Robust Against Domain and Channel Mismatches.
In SLT 2024.
(2024).
HAM-TTS: Hierarchical Acoustic Modeling for Token-Based Zero-Shot Text-to-Speech with Model and Data Scaling.
In ArXiv.
(2024).
Joint Speaker Encoder and Neural Back-end Model for Fully End-to-End Automatic Speaker Verification with Multiple Enrollment Utterances.
In Computer Speech & Language.
(2023).
CrossSinger: A Cross-Lingual Multi-Singer High-Fidelity Singing Voice Synthesizer Trained on Monolingual Singers.
In ASRU 2023.
(2023).
HiFi-WaveGAN: Generative Adversarial Network with Auxiliary Spectrogram-Phase Loss for High-Fidelity Singing Voice Generation.
In ISNN 2024.
(2023).
Improving Generalization Ability of Countermeasures for New Mismatch Scenario by Combining Multiple Advanced Regularization Terms.
In Interspeech 2023.
(2023).
Xiaoicesing 2: A High-Fidelity Singing Voice Synthesizer Based on Generative Adversarial Network.
In Interspeech 2023.
(2023).
Beyond Universal Transformer: block reusing with adaptor in Transformer for automatic speech recognition.
In ISNN 2024.
(2023).
Cross-Modal Audio-Visual Co-Learning for Text-Independent Speaker Verification.
In ICASSP 2023.
(2023).
SSI-Net: A Multi-Stage Speech Signal Improvement System for ICASSP 2023 SSI Challenge.
In ICASSP 2023.
(2022).
Deep Spectro-temporal Artifacts for Detecting Synthesized Speech.
In DDAM 2022 Workshop.
(2022).
Data Augmentation Using McAdams-Coefficient-Based Speaker Anonymization for Fake Audio Detection.
In Interspeech 2022.
(2022).
Spoofing-Aware Attention based ASV Back-end with Multiple Enrollment Utterances and a Sampling Strategy for the SASV Challenge 2022.
In Interspeech 2022.
(2022).
Attention Back-end for Automatic Speaker Verification with Multiple Enrollment Utterances.
In ICASSP 2022.
Skills
Languages and Tools
Python, C++, Shell, Git, MySQL
Deep Learning
PyTorch, PyTorch Lightning, Hugging Face
Speech Toolkits
SpeechBrain, WeNet, WeSpeaker, Kaldi, ESPnet
Generative Audio
Expressive TTS, codec and tokenizer design, voice LLMs
Multimodal AI
Audio-language modeling, full-duplex speech systems
Communication
Chinese, English, Japanese
Activities
Reviewer Service
- Conferences: NeurIPS, ICLR, ICML, ACL, ICASSP, ICME, INTERSPEECH
- Journals: IEEE OJSP, IEEE TASLP
Academic Activities
- ICASSP22, Interspeech22 Oral Presentation
- Organizing Committee, Joint Workshop of VoicePersonae and ASVspoof 2023
- SLT 2024, SVDD Challenge (Invited Talk): Shared insights on singing voice generation as a guest speaker
Open Source
- WeSpeaker contributor
- ASV-Subtools contributor
- HIVE dataset contributor
Experience
Independent Researcher
Independent
Conduct independent research on generative audio, voice LLMs, multimodal interaction, and intelligent audio understanding.
- Explore foundation models and data-centric methods for speech, audio, and multimodal AI
- Continue open research collaborations and publication work across generative audio and sound understanding
Senior AI Researcher
Led R&D of expressive TTS and full-stack voice-agent technologies for avatar and game products.
- Built KodamaTTS from scratch on a Qwen-based foundation model for virtual human and gaming applications
- Covered data curation, codec design, multi-node multi-GPU training, and evaluation in one pipeline
- Achieved sub-1 kbps codec quality with strong objective performance and reached top benchmark rankings in 3 languages
- Served as Co-PI on a joint project with Tsinghua University on cocktail-party speech interaction and audio separation
Multimodal Generative AI Researcher
Developed voice generation systems for Li Auto smart-space products and contributed to the multimodal foundation model MindGPT-4o.
- Proposed the GFSQ tokenizer for GPT-SoVITS to improve codebook utilization and decoding quality
- Trained a multi-timbre, multi-style voice generation model for in-car voice-blog scenarios in production
- Built synthetic-data workflows to scale accents, dialects, languages, emotions, and scenarios
- Led data production and audio-head pretraining and post-training for a full-duplex conversational model
Avatar Research Intern
Focused on high-fidelity 48kHz singing voice generation in collaboration with research and engineering teams.
- Upgraded XiaoiceSing to XiaoiceSing2 with adversarial training and achieved near-human MOS
- Developed HiFi-WaveGAN with a pulse-sequence design for stronger 48kHz singing synthesis quality
- Built CrossSinger for cross-lingual multi-singer SVS in English, Japanese, and Chinese
- Improved training efficiency with InstructSing and explored hierarchical acoustic modeling for voice LMs
Speech Recognition Researcher
Developed speech AI systems for Taobao Live compliance and broadcaster-risk control.
- Built a large-scale speaker recognition system for broadcaster identity verification in livestream scenarios
- Developed a spoken-term detection pipeline for policy-sensitive and illegal word monitoring
- Researched self-supervised speech representations and implemented an ESPnet-based end-to-end ASR system
