About the role
We are a small research lab building audio models for education, starting with teaching kids English.
We waited for frontier models to learn to teach kids effectively, but we waited in vain. Off-the-shelf models misunderstand kids' accented speech (word error rates up to 60% for younger children in our tests), can't pinpoint or correct pronunciation errors, fail to segment phonemes properly, can't match complexity to the learner's level, and the list goes on. Prompting and light fine-tuning don't fix these (we tried!).
That's why we're training state-of-the-art audio models, building on open-source foundations, on our proprietary corpus of millions of hours of English learning data. We already have working prototypes, and they're improving fast.
To go faster, we're hiring a senior research scientist to lead day-to-day training of our speech-to-speech model. You'll work with our overseas speech team and the researchers already fine-tuning our first prototype.
What you'll do
Drive the model. Build on our current prototype, evolve the data mix and training recipe for our full-duplex speech-to-speech model and cascaded systems, and take it to production.
Adapt it to children. Make it understand young, non-native English speakers and respond the way great teachers do: patient pacing, natural turn-taking and timely error correction.
Scale the training data. Extend our existing data pipeline, built with our overseas team.
Train at scale. Run and debug multi-GPU training, keep experiments reproducible, and get latency low enough for real-time conversation.
Measure what matters. Define evals for latency, turn-taking, child speech recognition and teaching quality, together with our education researchers.
Shape the research. Help set research direction, mentor engineers as the team grows, and publish what we learn.
What you'll bring
Hands-on speech model training. You have pre-trained or fine-tuned ASR, TTS or speech-to-speech models yourself, not just called APIs.
Conversational speech experience. You understand turn-taking, barge-in, backchanneling and streaming latency (ideally from work on full-duplex or spoken dialogue systems, but not essential).
Large-scale audio data. You have built pipelines over thousands of hours or more of real-world, messy audio.
Strong engineering. Fluent PyTorch, distributed training, and the patience to debug a diverging run at 2 a.m.
Research judgment. A PhD in speech, ML or EE, or equivalent industry experience training speech models, and a track record of setting direction on a project.
Nice to have
Experience with open speech language models such as Moshi, SpiritLM or GSLM.
Work on children's, accented or non-native speech.
Publications at Interspeech, ICASSP, NeurIPS, ICML or ICLR.
Widely used open-source speech code.
Experience at a speech lab or voice AI startup.
What we offer
Data no one else has. Millions of hours of authentic teacher–child conversations, plus fresh data from a live product.
Real ownership. A small team where your work shapes the whole model.
Compute. Dedicated GPU capacity for large-scale training.
Impact you can see. Your model teaches real children within months.
Competitive pay and meaningful equity.
Logistics
Location: on-site in Menlo Park, CA.
Process: an intro call, a technical deep-dive on a speech model you have built, a paid take-home or design exercise, and a meeting with the team. We aim to decide within two-three weeks.