<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>Scoutdm on PHP Boy Scout</title>
        <link>https://phpboyscout.uk/tags/scoutdm/</link>
        <description>Recent content in Scoutdm on PHP Boy Scout</description>
        <generator>Hugo -- gohugo.io</generator>
        <language>en-gb</language>
        <copyright>Matt Cockayne</copyright>
        <lastBuildDate>Fri, 20 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://phpboyscout.uk/tags/scoutdm/index.xml" rel="self" type="application/rss+xml" /><item>
            <title>Scout heard its first word</title>
            <link>https://phpboyscout.uk/scout-heard-its-first-word/</link>
            <pubDate>Fri, 20 Nov 2026 00:00:00 +0000</pubDate>
            <guid>https://phpboyscout.uk/scout-heard-its-first-word/</guid>
            <description>&lt;img src=&#34;https://phpboyscout.uk/scout-heard-its-first-word/cover-scout-heard-its-first-word_hu_ec54e699b9df25c.jpg&#34; alt=&#34;Featured image of post Scout heard its first word&#34; /&gt;&lt;p&gt;This is the first of a new kind of post, Trail Notes, which are short dispatches from building &lt;a class=&#34;link&#34; href=&#34;https://scoutdm.com/&#34;  target=&#34;_blank&#34; rel=&#34;noopener&#34;&#xA;    &gt;Scout&lt;/a&gt;, 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.&lt;/p&gt;&#xA;&lt;h2 id=&#34;sitting-in-the-room&#34;&gt;Sitting in the room&#xA;&lt;/h2&gt;&lt;p&gt;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&amp;rsquo;s voice channel, and the audio it&amp;rsquo;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.&lt;/p&gt;&#xA;&lt;p&gt;The first live run didn&amp;rsquo;t get as far as the room. The channel allowlist in the chat platform fails closed, which is what I&amp;rsquo;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&amp;hellip; and recorded absolutely nothing.&lt;/p&gt;&#xA;&lt;p&gt;The session asked, reasonably enough, whether I&amp;rsquo;d been in the channel. My reply the next morning was a sheepish &amp;ldquo;I was not in the channel during that run&amp;rdquo;.&lt;/p&gt;&#xA;&lt;p&gt;So the pipe was still unproven, and it had turned up something more useful: the manifest couldn&amp;rsquo;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&amp;rsquo;t attribute to anyone. The third run, with me in the voice channel this time (I checked), received 81 frames and kept 72.&lt;/p&gt;&#xA;&lt;h2 id=&#34;choosing-what-listens&#34;&gt;Choosing what listens&#xA;&lt;/h2&gt;&lt;p&gt;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.&lt;/p&gt;&#xA;&lt;p&gt;What I settled on for development is a small streaming model called kroko, run through sherpa-onnx. On a test corpus recorded with everyone&amp;rsquo;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&amp;rsquo;s conclusion, and I think it&amp;rsquo;s right, was that size mattered far less than what the model had been trained on. Its licence is murky, though, so it&amp;rsquo;s for development and evaluation only, with a licensed fallback waiting in the wings.&lt;/p&gt;&#xA;&lt;p&gt;A streaming model also needs a little silence padded on after the last frame before it&amp;rsquo;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&amp;rsquo;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&amp;rsquo;d go off and benchmark the wrong thing.&lt;/p&gt;&#xA;&lt;h2 id=&#34;phase-one-is-me&#34;&gt;Phase one is me&#xA;&lt;/h2&gt;&lt;p&gt;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&amp;rsquo;m the DM, and I didn&amp;rsquo;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.&lt;/p&gt;&#xA;&lt;p&gt;That&amp;rsquo;s the next dispatch. And for what it&amp;rsquo;s worth, the phrase I use to tell a session to go ahead and push is &amp;ldquo;the word&amp;rdquo;&amp;hellip; which I typed the same day Scout first kept anything it heard.&lt;/p&gt;</description>
        </item></channel>
</rss>
