About the job
You will join the Grok Voice Model team to help build the world’s best voice AI. We deliver smooth, natural, low-latency spoken interactions — expressive, multilingual, and reliable across devices and real-time scenarios. We own the full training pipeline: massive data curation, premium audio processing, frontier speech-language pre-training, and intensive post-training to push quality, speed, and stability to the limit. Our goal: make talking to AI feel like conversing with the most charming, kind, and knowledgeable person imaginable. We’re seeking exceptionally smart, execution-oriented engineers to help us get there.
Responsibilities
Design and execute large-scale speech data curation and processing pipelines, including collection of diverse real-world audio, synthetic data generation, and automated annotation workflows to enable high-quality model training and evaluation.
Work on pre-training and post-training of speech-language models, with targeted enhancements through supervised fine-tuning, reinforcement learning, and other techniques to ensure Grok Voice responses are accurate, factually grounded, natural and idiomatic in spoken style, conversational in tone, and fluent across multiple languages.
Build and iterate a comprehensive evaluation framework covering objective metrics (accuracy, quality, latency, expressiveness), human preference studies, content factuality assessments, real-time interaction quality, and experimentation infrastructure to measure and improve performance.
Work closely with product teams to integrate voice models into applications and real-time environments, define spoken interaction specifications, and handle the full lifecycle from prototype to global-scale deployment for stable, low-latency, delightful voice experiences.
Qualifications
Minimum
Python expert with deep proficiency in writing clean, efficient code for AI/ML systems.
Hands-on experience processing large-scale datasets using tools like Spark and Ray for cleaning, augmentation, and feature extraction.
Proficiency in pre-training and post-training speech-language models using JAX/PyTorch, including supervised fine-tuning, reinforcement learning, and optimizations for accuracy, factuality, natural spoken style, detail, and multilingual fluency.
Ability to set up and run rigorous evaluation pipelines: objective metrics, human preference studies, content factuality checks, and iterative A/B testing to drive model improvements.
Experience building or working with large-scale distributed training and inference systems on Kubernetes.
Proactive, self-driven attitude — ready to grind in a fast-paced, high-caliber team to deliver outstanding voice AI experiences.
Preferred
No preferred qualifications listed.