AI xPress

AI and computer science research.

Back to Foragentis

Speech & Audio

A new audio autoencoder encodes sound 10 times faster and uses 1.6 times less compressed code

Speech & Audio November 6, 2026

A hearing aid resting in an open palm

In a 2026 paper at the IEEE International Conference on Acoustics Speech and Signal Processing, researchers at Adobe Systems describe a new audio autoencoder. An autoencoder is a type of artificial neural network that compresses a signal into a small code and rebuilds the signal from that code. The authors report that their design encodes audio 10 times faster and uses 1.6 times less compressed code than existing approaches, while keeping reconstruction quality competitive. The design also raises the time compression factor from 2048 times to 3360 times.

Why does this matter for generative audio?

Generative audio models often rely on autoencoders to turn sound into a form the model can process. The authors say existing autoencoders are built reconstruction-first. That leads to high code rates, slow encoding, and separate models for different kinds of compressed representations and audio channel formats. The new design is generative-first. It handles continuous and discrete representations and common audio channel formats in one model. The authors say this could make generative modeling more tractable, because a 60-second single-channel signal compresses to 788 tokens. The abstract does not define token. In generative modeling generally, a token is a discrete unit the model processes.

How does the new approach perform?

The paper reports 10 times faster encoding, 1.6 times lower rates, and competitive reconstruction quality. It removes the need for channel-format-specific variants. The temporal downsampling factor rises from 2048 times to 3360 times. The authors say this balance of compression, quality, and speed reduces processing costs for applications that were previously constrained.

How could this affect speech recognition and speech synthesis?

The abstract does not report tests on speech recognition or speech synthesis. It describes an audio autoencoder, a building block that the authors say underpins generative models. Whether this version improves those later tasks is not stated in the abstract.

What do readers ask?

What is a latent rate?

The abstract does not define latent rate. As a general compression term, latent rate means how much compressed code a model keeps for a given amount of audio. The new model uses 1.6 times less compressed code than existing approaches, according to the paper.

What are tokens in this setting?

The abstract does not define token. In generative modeling generally, a token is a discrete unit the model processes. The paper says a 60-second single-channel signal compresses to 788 tokens.

Did the authors test this on speech systems?

The abstract does not say. It reports results for audio encoding speed, rate, and reconstruction quality, not for a finished speech system.

Limits and what's next

The abstract describes one model from researchers at Adobe Systems. It does not name the audio data or benchmarks used to measure the speed, rate, and reconstruction results. The results come from the authors' comparison with existing approaches, as described in the abstract, and the abstract reports no independent replication by another team. The abstract describes the reconstruction quality as competitive rather than improved. What remains unstated is how the model performs outside the authors' reported metrics.

Source: Jonah Casebeer et al., “A Generative-First Neural Audio Autoencoder,” IEEE International Conference on Acoustics Speech and Signal Processing, 2026.