Machine Learning Engineer, Voice Conversion at Cantina Labs

Role overview

This job description has been rewritten in fresh language using the facts in the listing. Source credit and original listing details are available below.

Cantina Labs is hiring a Machine Learning Engineer for its Speech Team to develop speech systems from data specification through production inference. The role focuses on voice conversion (VC) and related work such as controllable text-to-speech and voice design, connecting research, data, evaluation, and infrastructure to deliver reliable, cost-aware models.

You’ll build and train large-scale speech models, design experiments and quality evaluations, improve development tools, and contribute across the stack. The work also includes shaping data strategy, preparing training-to-inference pipelines for production, and supporting safeguards against misuse. The team values end-to-end ownership and close collaboration with research, data, infrastructure, and product.

The position is remote in the United States and full-time. Compensation is determined using factors including skills, experience, role scope, location, and market data. U.S.-based benefits include company equity; medical, dental, and vision premiums covered 99.99% by Cantina; 42 days of paid time off (15 PTO days, 10 sick days, 15 company holidays, and 2 floating holidays); parental leave and fertility support; a 401(k); a $500-per-month lifestyle spending account; and a One Medical membership. Complimentary lunch and snacks are provided for in-office employees.

Role at a glance: Build and ship large-scale voice-conversion and related speech models, from data and evaluation through production inference.

Responsibilities mentioned in the listing

• Architect, implement, pre-train, fine-tune, and post-train or align large-scale speech models, including methods such as GRPO and DPO.
• Design, run, and analyze experiments to improve understanding of speech models.
• Develop tools that improve team productivity and contribute across the stack, from low-level optimization to model design.
• Define data needs and collaborate on data acquisition, curation, augmentation, labeling quality, and synthetic-data strategies.
• Create objective and subjective evaluations, including listening tests, SV/WER/ASR-based metrics, robustness and bias checks, and red-team studies.
• Strengthen training, evaluation, and inference pipelines; profile latency, memory, and cost; and support production SLAs with monitoring and rollback.
• Contribute to safety and consent safeguards and mitigation of speech-technology misuse and abuse.

Qualifications stated in the listing

• Exceptional research or development experience with large-scale audio models exceeding 8B parameters and trained on more than 500,000 hours of data.
• Hands-on experience with diffusion and/or flow-matching transformers, including samplers, schedules, conditioning, and distillation.
• Hands-on experience training audio VAEs, neural audio codecs, and vocoders, including latent or tokenizer design, reconstruction and perceptual objectives, and adversarial training.
• Experience with multi-node, multi-GPU distributed training using FSDP, DeepSpeed, or equivalent.
• Strong software engineering skills and a track record of building complex systems.
• Proficiency with PyTorch and performance optimization, including profiling and CUDA, Triton, or C++ as needed, plus production-quality coding practices.
• Experience shipping large-scale speech/audio or multimodal generative models to production.
• Experience working with large-scale ML data and assessing quality with both subjective and objective signals.