Mind you, I’ve actually tried some of these newfangled dictation tools. Most of them are a simple transcription -> llm cleaning pipeline that doesn’t work nearly as well as they claim. You basically give up all control over the formatting/style. It’s also useless where precision matters and you can’t have random transcription errors lying around.
The companies behind it claim insanse things like a 10x speedup in wpm you can output. I tried it and it was… slower.
Part of the problem is me ig. Talking has always required effort for me, which seems to confuse other people. But when I’m dictating to one of these, I’m going slower (because talking clearly is hard) and I’m actively dumber (because talking takes brainpower).
That said, I did find them useful for when I can’t type/dirty hands/being lazy.
But there’s no need to pay an AI company $20/month for that. I run a model locally, get full privacy and no cost beyond a bit of battery
From what Ive heard Opencode, for agentic coding, is great, and I’d maybe try Gemma 4 (googles open weights model - permissive Apache 2.0 licensed) locally if I was so inclined. The size should be great for consumer hardware.
Mind you, I’ve actually tried some of these newfangled dictation tools. Most of them are a simple transcription -> llm cleaning pipeline that doesn’t work nearly as well as they claim. You basically give up all control over the formatting/style. It’s also useless where precision matters and you can’t have random transcription errors lying around.
The companies behind it claim insanse things like a 10x speedup in wpm you can output. I tried it and it was… slower.
Part of the problem is me ig. Talking has always required effort for me, which seems to confuse other people. But when I’m dictating to one of these, I’m going slower (because talking clearly is hard) and I’m actively dumber (because talking takes brainpower).
That said, I did find them useful for when I can’t type/dirty hands/being lazy.
But there’s no need to pay an AI company $20/month for that. I run a model locally, get full privacy and no cost beyond a bit of battery
What’s your model of choice? I want to try to get away from Claude and move it to a local model.
Claude? Claude doesn’t support audio input. I’m talking about transcription models. I use Parakeet v2
From what Ive heard Opencode, for agentic coding, is great, and I’d maybe try Gemma 4 (googles open weights model - permissive Apache 2.0 licensed) locally if I was so inclined. The size should be great for consumer hardware.
Thank you!