Voice used to turn any language into a clean English prompt. That pipeline is gone

Ideas Feedback
1

I used to be able to speak in whatever language I think in (Spanish, in my case). Cursor waited until I finished, including long pauses, and then produced a compact English summary of intent. It was not dictation. It was speech-to-prompt, always in English.

That is no longer what I get, and the desktop app and the web app fails with it.

How it used to work

- Speak in any language.

  • Pause as much as needed while thinking.
  • Get back a short, well-edited English prompt that described what I actually wanted.

That mattered for two reasons:

1. Many models follow instructions more reliably in English.
2. People who pause a lot while thinking never had those pauses baked into the text. The model waited, then summarized.

What happens now on desktop (macOS and Linux)

The Cursor desktop app (macOS and Linux; those are the ones I have tested) now live-dictates. It writes what I am saying as I say it, instead of waiting and summarizing what I meant.

If you talk in continuous, fluent English, that may be fine. If you pause, restart sentences, or think out loud, it is a mess. The pauses and half-thoughts land in the box, and I spend the next several minutes cleaning up text that the old system already cleaned up for me.

What happens now on the web

The web product is closer to the old behavior: it still summarizes instead of dictating. That part is still usable for people who pause.

The problem is the output language. I just confirmed this by using voice on the web: it generated a summary, not a transcript, but the summary came out in Spanish, because that is the language I spoke.

I do not want always a Spanish summary. I want the old contract: any language in, English summary out.

Ask

Please, think about restoring the previous pipeline, or at least make it selectable:

1. Summarize to English (legacy): speak in any language, tolerate pauses, emit a concise English prompt.
2. Summarize in spoken language: what the web appears to do now.
3. Dictate: live transcription, which is what the desktop app does now.

Until there is a way to get English summaries again, voice is worse for anyone who thinks in another language and still wants English prompts. On desktop it is also worse for anyone who pauses while thinking.

If there is already a setting to force English output, or to switch the desktop app back from dictation to summary, I would like to know.

Thanks!

5

Hey @ErunamoJAZZ, thanks for the post! Here’s the status of where things stand:

Voice never had a separate summarize step. The English you used to get came from the previous speech engine being pinned to English, so other languages came back translated. Voice now transcribes in the language you speak. On desktop it also writes as you go, instead of waiting for the whole take. The web app still waits until you finish, but the result follows the language you spoke. It is a transcription, not an English summary.

There is no setting to force English output, or to switch desktop back to waiting until you finish. The language menu next to the mic chooses the transcription language. Choosing English does not translate other languages into English.

In the meantime, a user rule such as “When I dictate in Spanish, treat it as instructions and work from a concise English reading of what I asked” can get you closer to that workflow.

6

Same here, but for thinking out loud rather than language.

I pause a lot, restart sentences, and speak at a different pace than live transcription expects. With the old wait-until-finished flow, that was fine. Cursor held the whole take, then gave me something I could send without much cleanup. I barely had to edit prompts.

Live streaming was how Cursor worked earlier, and I hated it. I would dictate in ChatGPT (or something similar), clean it up there, then paste into Cursor. Once Cursor added the wait-and-transcribe behavior, I stopped doing that and just used Cursor. Now that live dictation is back, I am editing again, and I am already tempted to go back to an external tool.

I am not asking to remove live transcription. It probably works well for people who speak in continuous, fluent takes. For speakers like me, it does not. If there is any way to make “wait until I finish” an option again, or otherwise support different speech paces and styles, that would help a lot. Right now it feels designed for one kind of speaker, and it does not fit everyone.