I work on neural audio codecs at LTCI, Télécom Paris, Institut polytechnique de Paris, with Xiaoyu Bie, Olivier Fercoq and Gaël Richard. My research focuses on discrete audio representations: how to quantize them, how to make their tokens interpretable (disentangled along frequency or signal components), and how to use them for tasks such as bandwidth extension.

The code of each paper lives in its own GitHub organization, which I created and own. Each one holds the code and a website with audio examples.

• Multi-rate BWE: bandwidth extension cast as token completion inside a frozen neural audio codec. A single transformer completes the high-band tokens for 8, 16, 24 and 32 kHz inputs, applied to both a spectral-domain codec (SpectroStream) and a waveform-domain codec (DAC).
• HP-codecX: a neural audio codec disentangled both across frequency (a low-rate branch and a high-rate branch coding the residual) and between harmonic and percussive components, whose tokens are used for bandwidth extension.
• Soft-Disentangled-Codec: a multi-branch codec where each branch runs at its own sampling rate and encodes the residual of the lower-rate branches, so that its tokens mainly carry the content of one frequency band.
• RandRVQ: a residual vector quantizer whose last stages use random, untrained codebooks drawn anew at every pass.

These projects reuse code from the Descript Audio Codec implementation.