This is the first of a new kind of post, Trail Notes, which are short dispatches from building Scout, my assistant for the Dungeon Master at a tabletop game. No grand argument, just what happened on the trail that week. This one goes back to the second week of September, when Scout got ears.
Sitting in the room
Until then a running Scout stayed out of the voice channels, since no table had asked it in yet. Between the 11th and the 13th that changed. A table configured for Scout is now activated when it starts, Scout joins that table’s voice channel, and the audio it’s allowed to keep lands as ordered segments with a manifest beside them. The manifest is written last, so it doubles as the fence that says nothing is still in flight.
The first live run didn’t get as far as the room. The channel allowlist in the chat platform fails closed, which is what I’d want it to do, right up until you notice that an empty list therefore permits nothing at all! One fix later, the second run joined, stayed in the channel for a full 58 seconds, left cleanly… and recorded absolutely nothing.
The session asked, reasonably enough, whether I’d been in the channel. My reply the next morning was a sheepish “I was not in the channel during that run”.
So the pipe was still unproven, and it had turned up something more useful: the manifest couldn’t tell the difference between a silent room and a room full of people it had refused to record. Now it records what it heard: how many frames came in, how many it kept, and how many it couldn’t attribute to anyone. The third run, with me in the voice channel this time (I checked), received 81 frames and kept 72.
Choosing what listens
Keeping the audio was only the start. Turning speech into words fast enough to be any use at a live table is a whole other kettle of fish, and that week also produced the recogniser, as a library ready to wire in.
What I settled on for development is a small streaming model called kroko, run through sherpa-onnx. On a test corpus recorded with everyone’s consent, it came out about level on word error rate with models many times its size, at a tiny fraction of the compute. The spike’s conclusion, and I think it’s right, was that size mattered far less than what the model had been trained on. Its licence is murky, though, so it’s for development and evaluation only, with a licensed fallback waiting in the wings.
A streaming model also needs a little silence padded on after the last frame before it’ll give up its last few words, and how much varies by model. So each model carries its own settings in a manifest, and Scout refuses to load one that doesn’t say how much padding it wants. Guess wrong and it drops the last word of every utterance, which looks for all the world like a worse model, and you’d go off and benchmark the wrong thing.
Phase one is me
The question that set up the next step was where any of this should be visible. My answer was that phase one is me. I’m the DM, and I didn’t want transcripts turning up in Discord. What I wanted was a small internal web page, run locally, where I could watch live transcription as it happened and judge it at a real table.
That’s the next dispatch. And for what it’s worth, the phrase I use to tell a session to go ahead and push is “the word”… which I typed the same day Scout first kept anything it heard.



