Updates

The current enclosure blocks outside noise well, but the small interior still holds a bit of residual reverberation — the phone's mic picks up a slightly echoey signal that doesn't quite match what a typical room sounds like. We're working on cutting that reverberation down so what the engine hears is closer to a normal listening environment.

The harness now runs on real Android and iOS phones placed inside a small acoustic isolation box, replacing the previous Android emulator setup. Playback speakers and the phone under test sit inside the box; an external computer drives the test and is wired to the phone. See the methodology page for the rig details. Legacy emulator runs are preserved but hidden from the public leaderboard so scores reflect only the new environment.

Semantic Error Rate (SER) and Character Error Rate (CER) now appear alongside WER on the run detail page. For Korean clips, SER and CER will likely replace WER as the primary metrics on the leaderboard, since they capture meaning-level and character-level accuracy better than WER's eojeol-tokenized scoring. English clips will continue to use WER.