I think the benchmark is misleading??

[marwanlhabti5-coder]

The multilingual quality is not completable and the performance is awful (there is no way a 3b model can use +21gb vram), and the English output is 24000khz
Thanks for open sourcing this tho but i think it need much improvement

[EvanCarlson]

Thanks for testing the model and sharing your feedback!

Could you share some examples of the multilingual issues you encountered? We’ve made improvements to VRAM usage (see here) and plan to benchmark potential bottlenecks soon. The model currently supports 24 kHz as it’s a research preview, which is controlled by the decoder. Training a new decoder at a higher sampling rate will improve this. We are planning to release our training code (including the decoder) in the near future!

We hope the open-source release encourages the community to help improve it. Contributions, bug reports, and optimizations are always welcome!