b8d91d9d90bb4231a7e4e437d0874db5Q: VoiceDash starts to substantially change what the user is dictating. Is this planned?
I've had multiple occasions, which have emerged recently, where what I am dictating is being changed. It seems you are allowing a lot more reinterpretation of what the user dictates.
For example, when I say,
“Please provide a summary of company A and include ABC,” it omits the first part of the sentence and starts to write the summary directly, including points ABC, as if I gave the instruction to write the summary, but I was only dictating and need the first sentence.
This happens if your prompt allows the LLM too much freedom to "improve" what the user dictates. I know since I build my own WisprSync before and know ehen this starts to happen.
You should really look into it. Unless this is what you are aiming for.
Please avise.
emad.puursina
Mar 25, 2026A: Thank you for reaching out.
I understand what you’re describing. Sometimes the AI can misinterpret the intent and treat a dictation as an instruction rather than transcribing it verbatim. This is not the intended behavior, and we are actively working to optimize our prompts to reduce these cases.
To help the AI process your dictation more accurately, you can use the custom prompt feature. Open the app, go to Settings, and find the custom prompt section. There you can add specific instructions to guide the AI to transcribe more literally and avoid reinterpreting your words.
Let me know if you need any help setting that up.
Thanks for the tip with the custom prompt. Did not see that. Great feature
I came to say the same thing. I asked for a to do list and just wanted to capture the things I need to do eg sort lead magnet and VD went on a full and unnecessary tangent about funnels.
If I set up a custom prompt as suggested will it always transcribe verbatim or is it more a case of me being able to instruct it to transcribe verbatim if I prompt "this is a to do list"?
Yeah, today it was full of this.
Not sure if Voicedash has changed the LLM. I developed my own version and the quality of the LLM makes a huge difference. You can use Qwen as a local LLM but the results differ greatly from commercial LLMs. Perhaps Voicedash has downgraded to a cheaper LLM?
This made the trick for me as CUSTOM PROMPT:
#HARD RULES:
Do not change the meaning of the original transcript text provided, except to:
- correct obvious grammatical errors and add punctuation, commas and insert paragraphs where it makes sense.
- insert lists when an obviously listing of terms was provided in the original transcript text provided.