About the Role

NEO is a home robot that handles chores and provides personalized assistance. It's meant to be controlled naturally: you talk, it understands, it responds, it acts.

As a Research Engineer, you'll own the speech and language stack that makes NEO feel like a calm, capable presence: streaming ASR, low-latency TTS, speech-to-speech interaction, and task-level NLU, all tuned for robotics constraints (latency, noise, far-field audio, on-device budgets, safety, and reliability).

In this role, you will:

• Build and deploy production speech models (ASR, TTS, and/or Speech to Speech) optimized for NEO's on-device compute and real-time latency requirements.

• Own the full pipeline from data collection and model training through deployment and monitoring.

• Solve hard acoustic problems that come with putting a robot in someone's living room: far-field recognition, noise robustness, barge-in handling, and multi-speaker environments.

• Design voice interaction that feels human: natural prosody, appropriate turn-taking, responses that match context and emotional tone.

• Build evaluation infrastructure, define quality metrics that actually correlate with user experience, and use real-world feedback to prioritize improvements.

• Collaborate with hardware and robotics teams to integrate speech capabilities with NEO's vision, memory, and physical behavior systems.

You might thrive in this role if you:

• Have 3+ years building speech systems, especially those currently running in production environments.

• Have deep expertise in at least one of: automatic speech recognition, text-to-speech synthesis, spoken language understanding, or speech-to-speech modeling.

• Are fluent in modern deep learning (transformers, diffusion models, autoregressive generation) and can train, debug, and optimize models end-to-end in PyTorch or JAX.

• Have deployed models under real constraints: on-device inference, latency budgets, memory limits, or edge hardware.

• Have published at top venues (ICASSP, Interspeech, NeurIPS, ICML) or built equivalent systems at companies known for voice AI.