About the role
Phone audio is hostile: 8kHz lines, accents, hold music, barking dogs, people talking over each other. Off-the-shelf speech models get us far — you'll get us the rest of the way, owning the research that makes Aurora hear and speak better than anything else on a phone call.
You'll sit between research and production: reading the papers, running the experiments, and shipping the winners into a pipeline that makes thousands of calls a day.
What you'll do
- Benchmark, fine-tune, and adapt STT/TTS models for telephone-quality audio across accents and noise conditions.
- Build the data flywheel — mining our own call corpus for training and evaluation sets.
- Prototype improvements in diarization, endpointing, and prosody that make conversations feel human.
- Partner with the voice-infrastructure team to ship research wins into production weekly.
What we're looking for
- Hands-on experience training or fine-tuning speech models (ASR/TTS) — not just calling APIs.
- Strong ML engineering: you own your data pipelines, training runs, and evals end to end.
- Publication record or equivalent shipped-model evidence in speech/audio ML.
- Pragmatism: you optimize for what ships, not what cites.
Nice to have
- Experience with low-resource or accent-robust ASR.
- Familiarity with streaming inference and on-line latency constraints.
- Multilingual speech modeling experience.