The field of neural speech and audio coding has witnessed an explosive growth, marking a paradigm shift from traditional signal processing approaches. Recent models have achieved unprecedented compression, delivering intelligible speech at bitrates often below 1 kbps.
However, most of these cutting-edge systems grapple with high computational complexity and considerable latency. This often renders them impractical for real-world telecommunication use cases, especially on resource-constrained devices like wearables, mobile phones, or IoT.
The inaugural 2025 Low Resources Audio Codec (LRAC) Challenge co-organized by Cisco tackles the above head-on, inviting engineers and researchers to design speech codecs that aren’t just low-bitrate, but also low-complexity, low-latency, and deployable on everyday hardware and in everyday communication environments. We are looking for solutions that can operate effectively at 1 kbps and 6 kbps bitrates in the real world.
This challenge runs from Aug. 1 to Sept.30 and culminates in a satellite workshop at ICASSP 2026, where participants will publish and present their solutions.
Challenge Focus: Low-Bitrate, Low-Complexity, Low-Latency, High-Impact
Unlike many research efforts that primarily push quality under ideal conditions, LRAC puts practical engineering first. This challenge tackles some of the most critical issues and opportunities of neural speech coding, demanding solutions that excel under strict constraints in:
- Bitrate: Pushing the boundaries of ultra-low compression, with target bitrates of 1 kbps and 6 kbps.
- Compute: Ensuring high computational efficiency for resource-constrained environments.
- Latency: Achieving real-time performance crucial for interactive applications.
Participants are encouraged to explore hybrid neural coding, real-time pipelines, and efficient quantization schemes—especially those suitable for embedded or CPU-only environments.
Related Work: Where LRAC Fits In
The journey towards neural codecs was significantly propelled by foundational models like WaveNet[6], which demonstrated the immense potential of neural networks for raw audio generation. Subsequent advancements like LPCNet[7] and SoundStream[8] built on those breakthroughs by applying neural network techniques to actual speech and audio coding, enabling real-time operation and high-quality reconstruction.
While these models and others have shown what’s possible, the quest for ultra-low bitrate and highly deployable codecs continues. The models listed below represent significant strides in achieving impressive compression, yet many still face the practical hurdles of high complexity and compute demands for edge deployment. The LRAC Challenge aims to bridge this gap by identifying and highlighting unresolved problem areas for the research community to address, with the goal of propelling the state-of-art towards truly practical solutions for resource-constrained devices.
| Codec | Bitrate | Notes |
|---|---|---|
| FocalCodec [1] | 0.16–0.65 kbp | Single binary codebook, semantic-aware, minimal compute |
| DualCodec [2] | 0.85–0.93 kbps | Semantic + waveform streams, open source |
| PSCodecDRLICT [3] |
~0.675 kbps | Prompt-based encoding, strong intelligibility |
| ESC [4] | ~1 kbps | Lightweight transformer with residual VQ |
| BigCodec [5] | ~1.04 kbps | High quality, but large (159M parameters) |
While these models set quality benchmarks, few are designed for real-time use on edge devices—a gap the LRAC Challenge aims to fill.
Who Should Participate?
- Speech and audio researchers working on compression, coding, or enhancement.
- ML engineers focusing on edge inference, streaming models, or low-power deployment.
- Anyone interested in pushing the state-of-the-art in codec design toward real-world applications.
Incentives for Engineers
- Rigorous benchmarking with comprehensive subjective testing battery.
- Opportunity to present and publish at the LRAC Workshop at ICASSP 2026.
- Join a vibrant community of engineers, researchers, and developers tackling neural audio and speech compression.
How to Get Started
If you’re ready to apply your expertise to a challenge with profound real-world impact and immediate relevance, join us for LRAC 2025 and help revolutionize speech technology. View the rules, including evaluation protocol, for more information.
Whether you already have a strong solution—or are simply curious to explore the problem space—we invite you to participate, connect with others, share ideas, and help advance the field together.
Sometimes, the biggest breakthroughs start with just trying.
References
[1] L. Della Libera, F. Paissan, C. Subakan, and M. Ravanelli, “FocalCodec: Low‑Bitrate Speech Coding via Focal Modulation Networks,” arXiv preprint arXiv:2502.04465, Feb. 2025
[2] J. Li, X. Lin, Z. Li, S. Huang, Y. Wang, C. Wang, Z. Zhan, and Z. Wu, “DualCodec: Dual‑Stream Neural Speech Codec with Semantic and Waveform Encoding,” Proc. Interspeech, 2025. Available: https://dualcodec.github.io
[3] Y. Pan, X. Zhang, Y. Yang, J. Yao, Y. Hu, J. Ye, H. Zhou, L. Ma, and J. Zhao, “PSCodec: A Series of High‑Fidelity Low‑Bitrate Neural Speech Codecs Leveraging Prompt Encoders,” arXiv preprint arXiv:2404.02702, Apr. 2024 (rev. Nov. 2024)
[4] Y. Gu and E. Diao, “ESC: Efficient Speech Coding with Cross‑Scale Residual Vector Quantized Transformers,” arXiv preprint arXiv:2404.19441, Apr. 2024
[5] D. Xin, X. Tan, S. Takamichi, and H. Saruwatari, “BigCodec: Pushing the Limits of Low‑Bitrate Neural Speech Codec,” arXiv preprint arXiv:2409.05377, Sept. 2024, and code at https://github.com/Aria-K-Alethia/BigCodec.
[6] van den Oord, A., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A., Kavukcuoglu, K. (2016) WaveNet: A Generative Model for Raw Audio. Proc. 9th ISCA Workshop on Speech Synthesis Workshop (SSW 9), 125
Arxiv: https://arxiv.org/abs/1609.03499
GitHub: https://github.com/huyouare/WaveNet-Theano
[7] J.-M. Valin and J. Skoglund, “LPCNet: Improving Neural Speech Synthesis Through Linear Prediction,” Proc. ICASSP, 2019
Arxiv: https://arxiv.org/abs/1810.11846
GitHub: https://github.com/xiph/LPCNet
[8] N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “SoundStream: An End-to-End Neural Audio Codec,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 495-507, 2022. doi: 10.1109/TASLP.2021.3129994
Arxiv: https://arxiv.org/abs/2107.03312
GitHub: no official implementation shared on GitHub.
The post LRAC Challenge 2025: Build Efficient Speech Codecs That Actually Work appeared first on Webex Blog.