Crazy about codecs
A site for comparing neural audio codecs. Various codecs are evaluated with the development set of VoxLingua107 across different languages. While these codecs are designed for clean, TTS data, VoxLingua107 is a in-the-wild dataset and is used here because it was interesting to compare the performance of various codecs under noisy conditions. Language specific-metrics are also reported for each codec.
| Codec | Bitrate | Param count | Mel L1 | PESQ | STOI | Speaker sim |
|---|---|---|---|---|---|---|
| bigcodec | 1.04 kbps | 159.4 M | 0.45 ± 0.11 | 1.96 ± 0.46 | 0.84 ± 0.09 | 0.77 ± 0.11 |
| facodec | 4.80 kbps | 29.0 M | 0.45 ± 0.06 | 2.09 ± 0.52 | 0.87 ± 0.08 | 0.79 ± 0.11 |
| longcat | 0.87 kbps | 709.8 M | 0.47 ± 0.05 | 1.84 ± 0.36 | 0.82 ± 0.09 | 0.66 ± 0.09 |
| mimi | 4.40 kbps | 79.3 M | 0.49 ± 0.05 | 3.09 ± 0.42 | 0.92 ± 0.05 | 0.91 ± 0.04 |
| nanocodec | 1.89 kbps | 62.6 M | 0.49 ± 0.05 | 2.37 ± 0.51 | 0.89 ± 0.07 | 0.86 ± 0.06 |
| neucodec | 0.80 kbps | 629.9 M | 0.49 ± 0.07 | 1.65 ± 0.29 | 0.78 ± 0.09 | 0.71 ± 0.11 |
| qwen3tts | 2.21 kbps | 170.6 M | 0.35 ± 0.04 | 2.58 ± 0.53 | 0.91 ± 0.06 | 0.93 ± 0.04 |
| snac | 0.98 kbps | 19.8 M | 0.52 ± 0.05 | 1.59 ± 0.27 | 0.79 ± 0.09 | 0.56 ± 0.11 |
Loading samples…
Other resources
- CodecMOS-Accent: A MOS Benchmark of Resynthesized and TTS Speech from Neural Codecs Across English AccentsWen-Chin Huang, Nicholas Sanders, Erica Cooper. arXiv:2603.14328, 2026.