Staff Research Engineer - Multimodal Generative Modelling

Synthesia · Full-time · Posted 1w ago

About the role

Synthesia's long-term vision is to build the best human-interactive models: systems that don't just talk to people, but perceive and respond to them, reacting to a user's actions and emotions, not only their words. Today over 60,000 businesses rely on our platform, and the next leap in what we can offer them depends on models that combine text, audio, and video into a single real-time interactive experience.

As a Staff Research Engineer, you'll join the Voice team within our 40+ person R&D department, but your scope will extend well beyond voice. You'll help define and drive that broader vision across teams, proposing ambitious research directions and taking direct ownership of the design and implementation of its most critical components. You'll work directly with our voice lead and collaborate tightly with our video teams and other senior members of the org.

Concretely, you'll work on voice to voice models that produce text and voice simultaneously. Models that can reason, interrupt the user with back channeling and talk. Models that feel like you are having a natural conversation with, without the feeling of turn taking that is dominant in current speech to speech models. Your role would be to partner in defining a roadmap, implementing it and shipping the outcomes to product.

What you'll do

Shape our roadmap to create new model capabilities and unlock new functionality for our customer base, on both short and long time horizons.

Propose novel multi-modal system architectures (especially text and voice).

Develop and evaluate streaming and conversational systems for low-latency, interactive voice-video synthesis.

Design solutions that reinforce emotional expressiveness and natural interaction.

Implement and bring designs to life, from pretraining through post-training.

Integrate and test novel architectures (neural codecs, diffusion, flow-matching) to enhance realism and responsiveness.

Define new evaluation metrics for conversational systems, including latency-aware and interaction-based measurements.

Track the latest research in audio-visual diffusion, autoregressive models, neural codecs, and multimodal LLMs.

Curate new datasets to complement existing data.

Lead post-training initiatives like DPO, fine-tuning, and distillation to bring models to shipping quality.

Ship models to production with optimised runtime to serve customers, and address their feedback thereafter.

You'll thrive in this role if you have

The ability to bring novel ideas and designs that advance the field of interactive multimodal systems.

Strong understanding of generative modelling, ideally applied to sequential or multimodal data.

Hands-on experience with large language models or similar transformer-based architectures.

High proficiency in PyTorch, including distributed training and model optimization.

A solid grasp of time-series modeling and tokenization, preferably in the context of audio, speech, or video.

A demonstrated ability to prototype quickly, test hypotheses, and iterate efficiently.

Proven experience training deep learning models end-to-end, from data preparation through evaluation.

Strong general software engineering skills, enabling contributions to a large, shared research infrastructure.

Particularly relevant experience

Having shipped a generative model into a live product used at meaningful scale, not just published or prototyped it.

Working on conversational or interactive systems where latency, responsiveness, and user experience were first-class constraints, not afterthoughts.

Working on LLMs with large scale trainings leadings to models with decent reasoning capabilities

Owning a research problem end to end: from architecture proposal through pretraining, post-training, and production deployment.

Collaborating across modalities or teams (e.g. audio and video, or research and product) to ship a unified system.

Bonus points for

Experience with real-time or streaming architectures.

Familiarity with state-of-the-art architectures in audio and speech generation, such as diffusion models, neural codecs, flow-matching models, or autoregressive decoders.

Excellence in one or more of the following modalities: voice, text, video.

Evidence of original research contributions, such as publications or open-source work at top-tier venues (e.g. NeurIPS, CVPR, ICML, ICLR, Interspeech).