LRAC 2.0 Challenge

Second edition of the Low-Resource Audio Codec (LRAC) Challenge · 2026

Registration is open Runs through Nov 17 — tap here to sign up.

Challenge is underway Runs Sep 1 through Nov 30 — a three-month sprint.

Registration is open Runs through Nov 17 — tap here to sign up.

Table of Contents

1. Call for Participation
2. Motivation
3. What’s New in LRAC 2.0

Call for Participation

We invite you to participate in LRAC 2.0 – the second edition of the Low-Resource Audio Codec (LRAC) Challenge, aimed at advancing the state of the art in neural audio coding for resource-constrained devices.

Building on the inaugural edition, LRAC 2.0 spans two tracks – Track 1: Transparency Codecs and Track 2: Memory-Efficient Transparency Codecs – and introduces three new directions: (i) propelling low-resource codecs toward transparency with increasing bitrate, (ii) rewarding memory-efficient solutions, and (iii) expanding language diversity across both training and evaluation.

LRAC 2.0 is a Signal Processing Grand Challenge and will be disseminated at ICASSP 2027 in Toronto.

• Call for Participation: LRAC-2.0-Call-for-Participation.pdf

• Registration: register here or see the Participate page for details.

• Questions: lrac-challenge@cisco.com.

Back to top

Motivation

The growing demand for speech interfaces on resource-constrained devices – including embedded systems, mobile devices, and general-purpose platforms with limited processing or power budgets – requires speech codecs that are efficient in terms of compute, memory, bitrate, and latency, while delivering high-quality speech under real-world conditions. Recent neural coding methods excel at optimizing these constraints individually; meeting them together, at deployment quality, remains an open challenge.

Back to top

What’s New in LRAC 2.0

Building on the first edition, LRAC 2.0 introduces three new themes:

• Bitrate scaling toward transparency. Beyond low and ultralow rates: can low-resource neural codecs keep improving with bitrate and reach transparency? Achieving transparency is what would make these codecs useful in practice for coding the main stream, in addition to redundancy coding. LRAC 2.0 provides the first common benchmark to find out.

• A dedicated memory-efficiency track. Compute counts alone don’t predict real-time behavior. The new Track 2 adds a hard parameter cap and a theoretical memory-bandwidth penalty, so leaner solutions rank higher at comparable quality and bitrate.

• Initial multilingual coverage. Training data expands in size and language diversity; evaluation extends beyond English (ENG) to Spanish (SPA) and Mandarin Chinese (CMN).

Back to top