Grain vs Bandwidth

Voice Signatures let Grain recognise speakers by voice rather than by audio channel, so transcripts label who said what even when several people share one mic or a capture comes from an unknown source. It ships alongside an upgraded transcription model, a split between workspace and account settings, fewer false positives from trackers, and a cleaner permission screen on shared meeting links. July's release put Grain inside ChatGPT and added MCP tools.

Having spent the spring and summer making Grain's data reachable from outside — MCP tools, bulk AI actions, an app inside ChatGPT — the work has turned to whether that data is worth reaching. Speaker attribution is the weak point in any transcript that did not come from a video call with one participant per stream, and Grain is explicit that better labels mean cleaner context for whatever model consumes them, its own or someone else's. In-person and local captures are the cases this unlocks, which is where meeting recorders built around a bot joining a call have the least to offer.

Expect voice identity to become a workspace-level asset rather than a per-meeting inference — recognising a colleague across every recording without re-labelling. How signatures are enrolled and what happens to them when someone leaves is not addressed in this entry.