Could you also set up persistent memory so it does this every time, or can it not tell when you're dictating? I'm thinking it still might be the latter, but man they've gotta fix that. I use memory for all kinds of stuff like that so I don't have to remember all my own weird rules!
For persistent memory: You can conceivably tell ChatGPT in the same chat, "I'm using voice dictation. Whenever I correct you, save the correction as a memory" - but this is chatbot-dependent in how they store memories, etc.
With agents, this is much easier, as they can save a Markdown file with your voice profile (which is actually exactly what the self-installing skill does) and it works regardless of which agent you use.
As for "tell when you're dictating," I experimented with this quite a bit when working on the agent skill, and originally had that as a feature to include. For instance, if the agent sees a long text with lots of rambling and "uh" and "uhm," it assumes I'm using voice input and automatically triggers the voice-mode skill. But in practice, it didn't trigger reliably, so you're still better off explicitly flagging the "I'm using voice now" if you want to be sure.
Thanks - following. I think we'll get there soon w/the silver platter (it'll know when you're using voice somehow). Til then, I will remain diligent. I liked this hack.
I'm working on a very similar problem right now because, at work, we have lots of Names and Acronyms (it's the Department of War after all) and I'm trying to build a customer name and acronym dictionary that a transcription agent would apply first.
The bigger issue is that MS teams does a halfway OK job separating speaking by person even if it gets the words wrong so I need the raw audio for what you describe and then the attepted transcript, and have it blend the two.
So it sounds like you need to catch the acronym/name-check at the point of transcription, rather than fixing it retroactively? That's more in line with how the third-party AI voice dictation tools work.
Substack Live also does a decent job of speaker identification, but not always 100%.
The agent never questions the input. That is the whole problem. I had a health endpoint answering 200 this week and I treated it as done. It was the previous build, answering perfectly.
A prompt that says pause and verify is still a request. The check that actually stopped me was the one that could refuse.
Could you also set up persistent memory so it does this every time, or can it not tell when you're dictating? I'm thinking it still might be the latter, but man they've gotta fix that. I use memory for all kinds of stuff like that so I don't have to remember all my own weird rules!
For persistent memory: You can conceivably tell ChatGPT in the same chat, "I'm using voice dictation. Whenever I correct you, save the correction as a memory" - but this is chatbot-dependent in how they store memories, etc.
With agents, this is much easier, as they can save a Markdown file with your voice profile (which is actually exactly what the self-installing skill does) and it works regardless of which agent you use.
As for "tell when you're dictating," I experimented with this quite a bit when working on the agent skill, and originally had that as a feature to include. For instance, if the agent sees a long text with lots of rambling and "uh" and "uhm," it assumes I'm using voice input and automatically triggers the voice-mode skill. But in practice, it didn't trigger reliably, so you're still better off explicitly flagging the "I'm using voice now" if you want to be sure.
Thanks - following. I think we'll get there soon w/the silver platter (it'll know when you're using voice somehow). Til then, I will remain diligent. I liked this hack.
Will this work for the voices in my head?
Only if you use the latest model, which I believe is "Brain 2.78"!
I'm working on a very similar problem right now because, at work, we have lots of Names and Acronyms (it's the Department of War after all) and I'm trying to build a customer name and acronym dictionary that a transcription agent would apply first.
The bigger issue is that MS teams does a halfway OK job separating speaking by person even if it gets the words wrong so I need the raw audio for what you describe and then the attepted transcript, and have it blend the two.
So it sounds like you need to catch the acronym/name-check at the point of transcription, rather than fixing it retroactively? That's more in line with how the third-party AI voice dictation tools work.
Substack Live also does a decent job of speaker identification, but not always 100%.
Would be curious how you end up solving it!
The agent never questions the input. That is the whole problem. I had a health endpoint answering 200 this week and I treated it as done. It was the previous build, answering perfectly.
A prompt that says pause and verify is still a request. The check that actually stopped me was the one that could refuse.