Deepgram
Deepgram
Deepgram je podjetje na področju glasovne umetne inteligence, ki razvija API-je za pretvorbo govora v besedilo, pretvorbo besedila v govor in razumevanje jezika. Njegova platforma razvijalcem pomaga ustvarjati glasovne aplikacije v realnem času, vključno s pogovorno umetno inteligenco, orodji za kontaktne centre in delovnimi postopki za medicinsko transkripcijo. Z obdelavo govorjenega jezika s poudarkom na natančnosti, hitrosti in stroškovni učinkovitosti Deepgram organizacijam v zdravstvu, logistiki, svetovanju in drugih sektorjih omogoča pretvarjanje glasovnih podatkov v koristne informacije in interaktivne izkušnje.

Staff Researcher, Voice AI Foundations

Develop latent-space models for Deepgram’s voice AI platform, advancing neural audio codecs, generative speech, multimodal systems, and efficient inference on hardware.

Opis

  • Develop neural audio codecs that combine very low bit rates with high-fidelity reconstruction.
  • Create steerable generative models for varied human speech, including emotional, multi-speaker, noisy, and overlapping speech.
  • Build embedding systems that separate latent representations into speaker, content, style, environment, and channel dimensions.
  • Recombine latent representations to generate synthetic audio data at unprecedented scale.
  • Train multimodal speech-to-speech systems for robust understanding and empathetic, human-like responses.
  • Design model architectures, training methods, and inference algorithms for bare-metal hardware.
  • Support cost-effective training on datasets spanning billions of hours and real-time inference for hundreds of millions of concurrent conversations.
  • Run controlled experiments, ablations, evaluations, stress tests, and edge-case analyses.
  • Advance foundational research in voice AI, speech, and language modeling.

Zahteve

  • Strong grounding in mathematics and statistical learning theory, especially self-supervised and multimodal learning.
  • Deep knowledge of foundation model architectures and scaling training across multiple modalities.
  • Ability to derive new mathematical formulations and implement them efficiently.
  • Demonstrated ability to build and curate large datasets while maintaining quality and diversity.
  • Proven experience designing controlled experiments to test architectural innovations and theoretical insights.
  • Experience preparing models for real-world deployment, including hardware constraints and efficiency methods.
  • Record of open-source contributions or research publications advancing speech or language AI.
  • Comfort using and experimenting with advanced AI tools.
  • Ability to identify decisive experiments that quickly validate or disprove ideas.
  • Vision for scaling successful proofs of concept by a factor of 100.
  • Experience or knowledge spanning neural audio codecs, generative models, embedding systems, latent representations, multimodal speech-to-speech systems, model architectures, training methods, inference algorithms, and hardware-efficient deployment.

Ugodnosti

  • Remote work arrangement.
  • AI-first environment that encourages using and experimenting with advanced AI tools.
  • Opportunity to pursue transformative voice AI research at scale.
  • Opportunities to contribute to open source and publish research.
  • The interview AI Notetaker recording may be declined without affecting candidacy.

Sorodna delovna mesta

AT&T

AT&T Business Sales Representative, Territory Accounts — Memphis

AT&T

Build AT&T’s Memphis small-business and mid-market customer base by selling fiber broadband and other communications solutions. The role combines prospecting, account development, and consultative selling to grow territory revenue.

Odpri