When a learner says a word in Sound Nativ, we tell them which sound was off and how to fix it, in about a second. That feedback is the product. If it is wrong, nothing else matters.
You will own the models and pipelines behind it: phoneme recognition, pronunciation scoring and accent analysis, from training run to production endpoint.
What you'll work on
- Training and fine-tuning speech models (wav2vec 2.0, Whisper-style encoders and friends) for phoneme-level scoring and accent detection.
- Deploying them behind our Python and FastAPI scoring service for real-time and batch inference.
- Building evaluation sets that cover the first languages our learners speak, and tracking how scores hold up across them.
- Data pipelines for audio and text: cleaning, labelling, augmentation and versioning.
- Exploring where LLMs and speech-to-speech models can coach, not just score.
Who you are
- Strong foundations in machine learning, statistics and signal processing, and pride in models that make it to production.
- Equal parts creative and methodical. You try the proven and the unproven, and you measure both.
- You care about the person on the other end. A score that feels unfair to a Korean speaker is a bug.
- You like long stretches of focused work without meetings, and you don't wait for permission.
Requirements
- 3+ years training, evaluating and deploying ML models in production, ideally in speech or audio.
- Fluent in Python and PyTorch.
- Experience with ASR, forced alignment or phoneme recognition.
- Comfortable owning infrastructure: GPUs, inference latency and cost.
Nice to have
- Background in phonetics, or a good ear for accents.
- Published research or open-source work in speech.
- Experience optimising models for low-latency or on-device inference.
What we offer
- Competitive pay, based on your experience.
- Work from anywhere.
- Versátil hours. We care about what you ship, not when you are online.
- Real ownership from week one, and a direct line to the founders.
- The AI tools you need to move fast, paid for.