Changelog

• New model, vocal8. Same size and CPU cost. It removes much more of a solo instrument playing on its own: trumpet 20.4 dB instead of 4.8 dB, flute 10.0 dB instead of 2.9 dB. Every voice test set is as good or better.
• Weaker in one case: singing buried under much louder music in poor audio keeps less of the voice than with vocal6. Tenor saxophone still mostly passes.
• The ON badge clears when the filtered tab closes or navigates away.
• The keyboard shortcut on another tab moves the filter to that tab.
• After ten filtering sessions the popup asks once for a review on the Chrome Web Store. The count stays on your computer.
• No new permissions. The extension still makes no network requests.

The macOS, Linux and Android apps ship the same model, so all four platforms report the same model hash.

• New model, vocal6, trained with more singing. It keeps singing much better: on singers it had not heard, the mean gain rose from +2.6 dB to +5.9 dB. Speech and songs also improved slightly.
• Same network size and CPU cost as before.
• Because the training data now includes GTSinger, the model weights are licensed CC BY-NC-SA 4.0. The credits are in CREDITS.txt.

• Switching modes crossfades instead of clicking.
• Starting a capture can be cancelled, and the icon’s badge only says ON while a tab is really being filtered.
• Internal engine fixes, and support for the newer model file format.
• Same model as 0.3.0.

• New model, vocal4, trained with instrumental tracks from the Free Music Archive. On music it had not heard before, it removes 2.6 to 4.7 dB more of it than the previous model.
• The trade: on songs from the MUSDB test set it gains up to 0.6 dB less than the previous model.