Research Scientist - Audio Codec

At Mirelo, we’re pushing the limits of what generative audio can do, and our ability to innovate depends heavily on the quality of our underlying audio representations. In this role, you’ll work at the core of our modeling stack - designing, training, and evaluating neural audio codecs that directly shape the performance of our next-generation music and sound models. You’ll collaborate closely with the model team, experiment with both continuous and discrete representations, and build the evaluation tools that help us understand what actually moves the needle. Your work will sit at the foundation of building the best-sounding generative models in the world.

Key Responsibilities

• Develop and implement new neural audio codecs for sound, music and speech that push the state-of-the art in sound quality and are optimized for the the use case of generative models.

• Think about the specific challenges that arise when the codec is primarily used as a latent representation in the context of generative audio models (in the end, the ultimate goal is to build the best audio generative models)

• Explore the trade-offs of continuous (as typically used for diffusion models) vs. discrete audio representations (as typically used for autoregressive models).

• Develop benchmarking pipelines for codec evaluation

• Conduct initial experiments with generative models to verify that a new candidate codec is actually useful for our downstream tasks.

Ideal Candidate Profile

• Strong background in deep learning for audio: neural codecs, source separation, speech models, or generative audio systems

• Specific hands-on experience in designing and training neural audio codecs

• Solid understanding of audio signal processing fundamentals

• Strong track record (research and/or open-source) in the field of audio ML

Nice to Have

• Hands-on experience with generative audio models and good intuition of how the choice of the codec influences the training and performance of the generative model

• Strong publication record (e.g., NeurIPS, ICML, ICLR, Interspeech, ICASSP, WASPAA)
  

Why Join?

• Join at a pivotal moment. We've secured fresh funding and are gaining traction - now is when your contributions can make a real difference to our success.