Machine Learning Engineer, Speech - Joint Audio-Video Modeling

Cantina Labs is a social AI company focused on developing advanced real-time models for expression and storytelling. They are seeking a Machine Learning Engineer to join their Speech Team, working on state-of-the-art speech and audio generation systems, particularly in joint audio-video modeling.

Responsibilities:

• Design, train, and improve the audio VAEs, neural codecs, and vocoders our generative models sit on top of latent design, reconstruction and perceptual objectives, compression-vs-fidelity tradeoffs
• Architect, implement, pre-train, fine-tune, and post-train/alignment (e.g., GRPO/DPO) diffusion and flow-matching transformers for large-scale audio and video generation
• Design the audio conditioning and cross-modal alignment inside joint AV models, audio latents alongside video latents, reference-audio and multi-speaker conditioning, multi shot generation audio/video modeling
• Design, run, and analyze scientific experiments to advance our understanding of the models
• Define data requirements and collaborate on acquisition, curation, AV-sync and quality filtering, annotation quality, and synthetic data strategies for paired audio-video and speech corpora
• Design automated objective/subjective evaluations audio fidelity and intelligibility metrics, AV-sync, listening and viewing tests, robustness & bias checks, and red-team studies
• Drive distillation, step-count reduction, quantization, and kernel/memory optimization to meet interactive latency and cost targets
• Harden the training → evaluation → inference pipeline; profile latency, memory, and cost; and meet production SLAs with robust monitoring and rollback
• Partner with infrastructure to run distributed training/inference on cloud fleets and productionize models with reliability and observability
• Independently lead small research projects while collaborating on larger team initiatives, including cross-team work with video generation
• Develop and improve dev tooling to enhance team productivity
• Contribute to safety/consent guardrails, watermarking, and misuse/abuse mitigation for responsible voice and likeness technology

Requirements:

• Exceptional research/development experience with large-scale audio models (>8B parameters, >500k hours of data)
• Deep hands-on experience with diffusion and/or flow-matching transformers, including practical knowledge of samplers, schedules, conditioning mechanisms, and distillation
• Deep hands-on experience training audio VAEs, neural audio codecs, and vocoders latent/tokenizer design, reconstruction and perceptual objectives, adversarial training
• Strong experience with multi-node, multi-GPU distributed training (FSDP/DeepSpeed or equivalent)
• Strong software engineering skills with a proven track record of building complex systems
• Strong with PyTorch and performance work (profiling, CUDA/Triton/C++ as needed) and writing reliable production-quality code
• Shipped large-scale speech/audio or multimodal generative models to production
• Background in working with large-scale ML data, and the ability to iterate on data and triangulate quality using both subjective and objective signals
• Experience with voice cloning, speech control/steerability, or expressive speech generation
• Notable publications and/or open-source contributions in speech/audio/ML
• Experience with multimodal audio-video modeling: joint AV generation of multi-shot, multi-speaker scenes with dialogue, music, and sound design generated jointly with video, and the cross-modal alignment that keeps them in sync
• Experience with video generation: video diffusion/flow-matching transformers, video VAEs, conditioned and multi-shot generation, building data pipelines for video models
• Streaming or real-time generation, causal distillation (e.g., Self Forcing / Self Forcing++)