I’m a Senior Staff Research Scientist
working on speech and conversational AI.
I’m at Meta, Super-Intelligence Lab, where I work on natural, expressive voice experiences and the systems that bring them to life. My work spans conversational speech, multilingual text-to-speech, and efficient inference across servers and devices.
I joined Meta in April 2017. I earned my Ph.D. and M.S. in Electrical Engineering and Computer Science from MIT, and my B.S. in Electrical Engineering from the University of Waterloo.
Building natural, expressive voice experiences—and the systems that make them possible.
Muse Realtime Avatar brings characters to life in live conversations, synchronizing speech with expressive facial and full-body animation.
My contribution I drove improvements in speech naturalness and expressiveness, helping Muse’s avatars speak with a more lifelike, engaging voice.
The Meta AI app introduced a full-duplex voice demo that generates speech directly from conversational dialogue, enabling more natural spoken interactions.
My contribution I helped develop full-duplex voice technology for Meta AI, shaping a more fluid, spontaneous experience that feels closer to human conversation.
Voice experiences across Ray-Ban Meta glasses, Oculus/Quest, and Portal—connecting smart glasses, mixed reality, and home devices through speech.
My contribution I led development of text-to-speech voices for Meta’s hardware devices, spanning server and on-device deployment. The work brought together multilingual support and a broad range of voice capabilities across Ray-Ban Meta glasses, Oculus, and Portal.
AI-powered tools for expression and interaction: Instagram text-to-speech and voice effects, custom characters in AI Studio, and voice-enabled apps built with the Voice SDK.
My contribution I partnered with product teams to bring AI capabilities into creative tools, supporting experiences across Instagram, AI Studio, and the Voice SDK.
Model and runtime optimization for real-time speech synthesis, from efficient CPU-based serving to lightweight mobile deployment.
My contribution I worked extensively on inference optimization across server GPUs and mobile edge devices, making high-quality models practical within the compute and latency constraints of real products.
Product imagery: Meta · Inference illustration: AI-generated
Recent work on speech, language, and the models that connect them. Full list on Google Scholar
SLT 2026 · Accepted
An asynchronous speech framework lets a model share short progress updates while it continues reasoning. Coordinating a reasoning stream with a lightweight speaking stream reduces audible silence while preserving answer accuracy in spoken reasoning and question answering.
arXiv · Preprint
A speech reward model evaluates naturalness through acoustic features and explicit reasoning, making its judgments easier to interpret. Trained with expert feedback, it aligns closely with human ratings and can guide reinforcement learning toward more natural speech.
ICLR 2026
SiTok combines semantic representation learning with diffusion-based audio reconstruction. Scaling the tokenizer improves speech understanding, reconstruction, and generation while operating at just 12.5 tokens per second and 200 bits per second.
ICLR 2026
VowelPrompt translates vowel-level pitch, energy, and duration into natural-language cues. These cues help language models recognize emotion using both the words and how they are spoken, with explanations grounded in prosody and context.
Speech Synthesis Workshop 2025
SemAlignVC aligns text and audio representations to separate linguistic content from speaker identity. This reduces leakage of the original speaker’s timbre during zero-shot voice conversion, improving target-voice similarity, intelligibility, and naturalness.
ICASSP 2025
TTS-Llama adapts a language model for speech synthesis. MoLE-Llama then combines text and speech through parameter-efficient late fusion and a mixture-of-experts design, supporting spoken responses while retaining text question-answering capabilities.
arXiv · Technical report
A technical report on the Llama 3 family of multilingual models for coding, reasoning, and tool use. It describes training and evaluation at scale, including a 405-billion-parameter model, and explores adding image, video, and speech capabilities.
ICASSP 2024
A hybrid vocoder jointly optimizes an acoustic model with differentiable signal processing. It produces speech with quality comparable to neural vocoders while using very little computation, making efficient speech synthesis more practical for constrained devices.
ICASSP 2024
A shared model handles text normalization, part-of-speech tagging, and ambiguous-word pronunciation together. Combining shared representations with task-specific heads improves overall performance, while a balanced homograph dataset strengthens pronunciation disambiguation.
Research papers and preprints from the past three years · Updated October 2026
For research conversations, collaborations,
or just to say hello.