Principal Research Scientist Speech Voice Foundation Models
Job Description
Principal Research Scientist Speech & Audio Foundation Models
\n
Text-to-speech, voice cloning, speech synthesis, realtime conversational voice.
\n
\n
$270,000–$500,000 base plus bonus, equity and benefits (US).
\n
Relocation assistance available. Visa transfer supported
\n
San Francisco on-site preferred | Remote considered in the US, UK and parts of Europe.
\n
Permanent, full-time.
\n
\n
\n
A top end research lab building realtime voice models text-to-speech, speech-to-text and speech-to-speech delivered as an API. The models run in production behind consumer applications used at very large scale, across health, learning, therapy, companionship, media and gaming.
\n
\n
Text-to-speech, voice cloning, speech synthesis, realtime conversational voice.
\n
\n
\n
The role
\n
\n
Build the models that are the product! This is full-stack research ownership: you frame the question, run the experiments, and ship the result. Research is only finished when it is in production and measurable.
\n
\n
\n
Responsibilities
\n
\n
\n
- \n
- Train foundation models: pre-training, reinforcement learning, reward modelling, post-training, new architectures, scaling.
- Build and improve voice and speech models across TTS, STT and speech-to-speech.
- Design the evaluation that proves the work: benchmarks, eval loops, LLM-as-judge, failure analysis. Evaluation is treated as a research product in its own right, not as a pre-launch checkbox.
- Work on frontier problems adjacent to the roadmap: multimodal, agents and tool use, test-time compute.
- Take models into production alongside the serving engineering team, inside a sub-200ms latency budget and across 100+ languages.
\n
\n
\n
\n
\n
\n
\n
\n
Essential
\n
\n
\n
- \n
- Hands-on foundation-model training. Pre-training, RL, reward modelling, post-training, scaling. Fine-tuning or building on top of someone else's model is a different discipline and is not what this role is.
- Real voice or speech research: TTS, STT or speech-to-speech. Speech-to-speech is the strongest signal; TTS and ASR both count. Text-only research does not transfer.
- Evidence you can point at: papers, shipped models, open-source contributions, or systems in production.
\n
\n
\n
\n
\n
\n
Desirable
\n
\n
\n
- \n
- Evaluation depth: benchmarks, eval loops, quality measurement, failure analysis.
- Publications at ICML, ICLR, NeurIPS, EMNLP, ACL, AAAI, Interspeech or ICASSP.
- PhD in ML or NLP, or equivalent practical experience you can point to.
- Frontier exposure: multimodal, agents, tool use, test-time compute.
- Public work: side projects, open-source, technical write-ups.
\n
\n
\n
\n
\n
\n