Skip to main content

Live Transcription with Inline Editing

Feature Request

Firstly, this product is awesome so thanks for making it.

Here’s one issue I would love to resolve: I’m using text to speech and try to say “feature-2.md” but end up with “feature2.md” or “feature dash 2.md” etc. There are many examples like this and in most cases there simply isn’t enough context in what I’m saying to know what I want as the final text.

Idea: when using live transcription (e.g. with soniox) I want to see what I’m saying as I am speaking and I want to be able to edit as I go. So if I see a typo I can just fix it right there and if I’m about to try to say something “~/.gitconfig” which I’m sure will be transcribed incorrectly I can just type it myself and then continue to speak.

Another idea could come out of this: you log the corrections I make and use that somehow to improve future transcriptions.

Posted by Oliver Chambers ·
4 comments

Log in to comment and vote

Comments4

  • Vowen

    Team•

    Mar 18

    Thanks @Oliver Chambers !

    We’re working on recognizing common phrases used while transcribing, like “dash”, “new line”, “comma”, “semi-colon”, “slash” etc. I think that should help a bit, but based on the context, that too can have a margin of error. However, in cases where there is no ambiguity, it should get it right.

    Regarding “ability to edit live transcription”, we’ll brainstorm how best to implement this with the current setup we have, and get back to you. It will be really helpful if you could share a couple more examples when you come across them, where it would have made it easier for you work with it, that will further give us the right idea as to how to best implement this.

    Finally, regarding capturing corrections, we’re considering a way to share feedback on your transcriptions from within the app, and that will be fed as context into the transcription engine for future transcriptions. That will improve / tailor the result according to your expectations. This should be added in the coming update to the app along with recognizing common phrases.

    In the meanwhile, I can suggest adding “custom instructions” to “Enhance transcription with AI” setting under “Configure your AI” page. That will help keep the result of transcription in line. You could add something like:

    Infer if the user is referring to a file from the input and make sure they are properly formatted.

Examples:
Input: ...feature dash 2.md....
Output: ...feature-2.md...

    The models we’ve noticed to work really well are Gemini Flash 2.5 Lite and Groq Llama 70B.

    • Oliver Chambers

      •

      Mar 18

      Thanks! Mostly I am using vowen for talking to claude code, so my examples mostly involve symbol names from source code.

      If I say “profile management panel” I could mean

      • profileManagementPanel

      • profile-management-panel

      • profile_managemenet_panel

      It depends on the language etc. and It’s not practical to add every single symbol name.

      That why I feel that it would be easier to just type than say “profile underscore management underscore panel” all the time.

      It is particularly bad for short context-specific words. For instance, I might say “extend generated files with .g” I will often get various different interpretations of “.g” from “dot g” or “dotty” or whatever.

      I am using Groq Llama and it is impressively good, but a seamless way to handle these tricky edge cases would be awesome.

      I suppose with live TTS I could see the mistakes as they happen and add a rule which says “when I say CORRECTION then the last sentence was transcribed incorrectly and fixup instructions will follow” which might help a bit too.

      • Vowen

        Team•

        Mar 18

        Are you using it mainly for tagging files in Claude code? We have an experimental feature that allows File tagging in Cursor / Windsurf and Anitgravity. A similar mechanism will work for Claude code as well. That’s the part we’re trying to figure out, because we can’t know beforehand if what’s running in a terminal is a Claude code session or not.

        If it’s mainly for tagging files, then we’ll find a way to make it work. But for variable names, we might have to rethink how to make it more reliable.

        We’ll brainstorm on the correction part as well.

        • Oliver Chambers

          •

          Mar 18

          not really “mainly” tagging files, but files paths are definitely a pain point.