Yang Yang

I am a Staff Software Engineer at Google based in San Diego, specializing in the intersection of generative AI and on-device deployment. My current work focuses on neural speech synthesis, enhancement, and audio codecs, and I previously worked on neural video compression. I am passionate about the full lifecycle of AI—from generative modeling research to the engineering rigor required to bring features into production.

Prior to Google, I spent several years at Qualcomm in both the Wireless R&D (patents) and AI Research divisions. I hold a Ph.D. in Wireless Networking from The Ohio State University (2015) and a Bachelor’s degree from Shanghai Jiao Tong University (2009).

selected publications

arXiv

DiffSoundStream: Efficient Speech Tokenization via Diffusion Decoding

Yang Yang, Yunpeng Li, George Sung, Shao-Fu Shih, Craig Dooley, Alessio Centazzo, and Ramanan Rajeswaran

arXiv preprint arXiv:2506.22362 2025

TLDR HTML

DiffSoundStream halves the required token rate of a neural speech codec by conditioning acoustic tokens on semantic tokens to reduce redundancy, then using a distilled latent diffusion decoder to reconstruct high-quality waveforms from semantic and coarse acoustic tokens in as few as four steps.
ICASSP

Binaural Angular Separation Network

Yang Yang, George Sung, Shao-Fu Shih, Hakan Erdogan, Chehung Lee, and Matthias Grundmann

In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2024

TLDR HTML Google Research Blog

A two-microphone speech separation model trained entirely on simulated RIRs that exploits TDOA "delay contrast" for angular source separation, generalizing across device geometries and running in real time on-device.
ICASSP

StreamVC: Real-Time Low-Latency Voice Conversion

Yang Yang, Yury Kartynnik, Yunpeng Li, Jiuqiang Tang, Xing Li, George Sung, and Matthias Grundmann

In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2024

TLDR HTML Poster Samples

StreamVC converts voice timbre in real time on mobile devices by combining causal soft speech units, whitened pitch features for speaker-agnostic prosody, and the SoundStream codec architecture for low-latency, high-quality synthesis.
ICASSP

Guided Speech Enhancement Network

Yang Yang, Shao-Fu Shih, Hakan Erdogan, Jamie Menjay Lin, Chehung Lee, Yunpeng Li, George Sung, and Matthias Grundmann

In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2023

TLDR HTML

GSENet takes both raw microphone signals and beamformer output as input, using a contrastive training scheme to let the neural network learn spatial rejection cues from the beamformer while remaining a transferable standalone module across device form factors.
NeurIPS

Neural Topological Ordering for Computation Graphs

Mukul Gagrani, Corrado Rainone, Yang Yang, Harris Teague, Wonseok Jeon, Roberto Bondesan, Herke Hoof, Christopher Lott, Weiliang Zeng, and Piero Zappi

In Advances in Neural Information Processing Systems (NeurIPS) 2022

TLDR HTML

Topoformer, an attention-based GNN encoder paired with a priority-based decoder, learns to minimize memory usage during computation graph scheduling by generating near-optimal topological orderings of DAGs, outperforming heuristic baselines on both synthetic and real-world compiler graphs.
ICLR

Transformer-based Transform Coding

Yinhao Zhu*, Yang Yang*, and Taco Cohen

In International Conference on Learning Representations (ICLR) 2022

TLDR HTML

A swin-transformer based auto-encoder structure is proposed that achieves SOTA in rate-distortion-complexity trade-off of image compression. It also outperforms SSF in video compression. AFAIK SwinT-ChARM is the first neural image codec that outperforms VTM in rate-distortion while with comparable decoding time on GPU.
ICIP

Progressive Neural Image Compression With Nested Quantization And Latent Ordering

Yadong Lu*, Yinhao Zhu*, Yang Yang*, Amir Said, and Taco S Cohen

In IEEE International Conference on Image Processing (ICIP) 2021

TLDR HTML

A prefix-decodable bitstream (lower bitrate stream is embedded as a prefix of higher bitrate stream) is obtained with nested quantization and per-element sorting by prior stddev, based on the hyperprior model.
ICASSP

Feedback Recurrent Autoencoder

Yang Yang, Guillaume Sautière, J. Jon Ryu, and Taco S Cohen

In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2020

TLDR HTML

Extend VAE to sequential setting for efficient representation learning of sequential data. In the application of speech compression, when paired with a neural vocoder, the proposed FRAE signifantly outperforms Opus, an open source speech codec.
ACCV

Feedback Recurrent Autoencoder for Video Compression

Adam Golinski*, Reza Pourreza*, Yang Yang*, Guillaume Sautière, and Taco S Cohen

In Asian Conference on Computer Vision (ACCV) 2020

TLDR HTML

One of the early works that propose neural network based video codecs. A recurrent layer is introduced in the decoder to capture long-range redundancy of video content, the state of which is fed back to encoder.
CVPR

Guided Variational Autoencoder for Disentanglement Learning

Zheng Ding*, Yifan Xu*, Weijian Xu, Gaurav Parmar, Yang Yang, Max Welling, and Zhuowen Tu

In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2020

TLDR HTML

An auxiliary decoder is used to improve the disentanglement of latent from a VAE.