Research
WideCodec matches NVIDIA NeMo-44k on naturalness at a tenth of the bitrate
A single codebook at 50 tokens per second, scored against four other 44 kHz-class codecs on a 9,291-clip multilingual wideband set.
Research from Scicom AI Enterprise Solutions has improved WideCodec over our 2026-07-21 release, UTMOSv2 from 2.788 to 2.919 on a multilingual wideband benchmark matching NVIDIA NeMo-44k (2.903) at roughly 1/10 of its bitrate (0.8 kbps vs 8 kbps, single codebook, 50 tokens/s).
UTMOSv2 naturalness against bitrateHigher is better. Further left is cheaper.
Show the data
| System | Bitrate (kbps) | UTMOSv2 naturalness |
|---|---|---|
| WideCodec 2026-08-16 (Scicom) | 0.8 | 2.919 |
| WideCodec 2026-07-21 (Scicom) | 0.8 | 2.788 |
| NVIDIA NeMo-44k (NVIDIA) | 8.0 | 2.903 |
| dac | 8.0 | 2.672 |
| snac44k | 2.6 | 2.340 |
| encodec48k | 24.0 | 2.042 |
The gain over July is +0.131 UTMOSv2 at the same character error rate. Ground truth on this set sits at 2.822, which both the August release and NeMo-44k now pass.
What WideCodec is
WideCodec is our speech tokenizer, designed to deliver maximum compression with minimal compromise:
- 44 kHz high-fidelity audio
- Pareto-best quality against bitrate, as the chart above shows
- Benchmarked across 400+ languages
- A single codebook at only 50 tokens per second
- LLM-ready, making it significantly easier to post-train language models into speech models without modifying the model architecture
Smaller token streams mean more efficient training and inference, with strong CER recovery across multilingual evaluation. This is a step toward making high-quality speech AI more efficient, scalable and accessible, from research to production.
The repository is on Hugging Face: Scicom-intl/WideCodec.
First published on LinkedIn in two posts: the release on 5 August 2026, and the results on 19 August 2026 by Husein Zolkepli, VP Research. The figures are reproduced as published.