SpeechCompass (Best Paper Award @ ACM CHI 2025, paper + video in repo) addresses this by showing the direction of each speaker. Captions are color-coded with an arrow pointing toward where the speech is coming from. We show it on phones with 2+ mics (180°), and release the firmware and hardware for a 4-mic phone case for full 360°.
It uses fast time-difference-of-arrival algorithms that can run on small, low-power microcontrollers. It is private by design and runs fully on-device: it uses spatial direction, not voiceprints, and no audio leaves the phone (for direction). The repo has the Android app (prebuilt APK in Releases), the portable C11 DSP library, firmware, and PCB designs, all Apache 2.0. Happy to answer questions and interested in feedback, especially from folks who use captioning daily.