Remote opportunity at
DeepgramResearch Staff, Voice AI Foundations [UK - Remote]
Deepgram offers real-time speech-to-text and text-to-speech APIs and tools for building voice agents, which customers can run in its cloud or install on their own infrastructure.…
Career Tools
About This Role
Deepgram offers real-time speech-to-text and text-to-speech APIs and tools for building voice agents, which customers can run in its cloud or install on their own infrastructure. The company is recruiting a Research Staff member for its Voice AI Foundations team, a remote role based in the UK. The researcher will develop Latent Space Models, an approach meant to tackle the data, scale and cost limits that hold back voice AI.…
Job Description
Deepgram offers real-time speech-to-text and text-to-speech APIs and tools for building voice agents, which customers can run in its cloud or install on their own infrastructure. The company is recruiting a Research Staff member for its Voice AI Foundations team, a remote role based in the UK.
The researcher will develop Latent Space Models, an approach meant to tackle the data, scale and cost limits that hold back voice AI. Topics named in the posting include compact neural audio codecs with strong fidelity, steerable generative models that produce a wide range of human speech, and embedding systems that split the codec latent space into controllable factors such as who is speaking, what is said, delivery style, room acoustics and recording channel. The work also covers synthetic audio generation at large scale through latent recombination, multimodal speech-to-speech systems, and model architectures and training methods suited to real hardware.
The posting asks for a solid mathematical base in statistical learning theory, especially for self-supervised and multimodal learning, along with deep knowledge of foundation model designs and how to scale them across modalities. Candidates should be able to derive new formulations and implement them efficiently, build pipelines that curate very large datasets while keeping quality and diversity high, and design controlled experiments that isolate architectural effects. Experience optimizing models for deployment is also listed.
Deepgram describes an AI-first culture in which staff are expected to use and experiment with AI tools in daily work. The company says the pace of change is fast and that the role may not suit people who want a highly prescriptive, traditional schedule. No salary range is given in the posting.
Responsibilities
- Develop Latent Space Models for robust, context-aware voice AI
- Build neural audio codecs that compress audio heavily while keeping high fidelity
- Create steerable generative models that synthesize diverse speech, including noisy and multi-speaker scenes
- Design embedding systems that split the codec latent space into controllable dimensions
- Apply latent recombination to produce large volumes of synthetic audio
- Train speech-to-speech models that combine several modalities and reply with empathy and purpose
- Design model architectures and inference methods suited to hardware efficiency
Requirements
- Solid mathematical background in statistical learning, especially self-supervised and multimodal methods
- Advanced command of large foundation models and how to scale training across modalities
- Able to turn new mathematical ideas into efficient code
- Past work on data pipelines that process very large corpora with consistent quality
- Record of designing clean experiments that separate the effect of each architectural change
- Know-how in tuning models for production use under hardware limits
Qualifications
- Public code or papers that moved speech or language AI forward
Core Skills
Frequently Asked Questions
Answers are based only on the employer’s listing; where it doesn’t say, neither do we.
Is the role remote?
Yes. The job title specifies a UK-based remote role.
Is a salary stated?
No. The posting gives no salary or pay range.
Is AI tool use expected?
Yes. Staff are expected to use and try out advanced AI tools as part of daily work.
What does the research cover?
The posting names neural audio codecs, steerable generative speech models, latent-space embeddings, synthetic audio generation and multimodal speech-to-speech systems.
Sample Interview Questions
AI-generated questions tailored to this specific role — a preview of the full practice set.