Disentangling speech from surroundings with neural embeddings

TL;DR

This paper introduces a neural embedding-based method to effectively disentangle speech from environmental noise and reverberation, enabling cleaner audio separation and targeted audio adjustments.

Contribution

A novel training procedure for neural audio codecs that produces structured embeddings, separating speech from environmental factors in the embedding space.

Findings

Successful separation of speech from noise and reverberation

Ability to modify audio characteristics through embedding manipulation

Structured embeddings enable targeted audio editing

Abstract

We present a method to separate speech signals from noisy environments in the embedding space of a neural audio codec. We introduce a new training procedure that allows our model to produce structured encodings of audio waveforms given by embedding vectors, where one part of the embedding vector represents the speech signal, and the rest represent the environment. We achieve this by partitioning the embeddings of different input waveforms and training the model to faithfully reconstruct audio from mixed partitions, thereby ensuring each partition encodes a separate audio attribute. As use cases, we demonstrate the separation of speech from background noise or from reverberation characteristics. Our method also allows for targeted adjustments of the audio output characteristics.