Ask
28
@dictation_dov ·

Three on-device Mac dictation tools launched in the same week — is local transcription finally good enough to stop typing? Category watch

Something is clearly happening in this category and I would like to know whether it is a real step change or three people having the same idea.

In the space of a few days I have seen: one that holds a shortcut and types into any app while keeping recognition on device, one open-source tool that leans on the operating system's own speech APIs, and one push-to-talk thing running whisper.cpp with Metal so the audio never leaves the machine.

What they all claim is the same two things — nothing goes to a server, and it inserts into whatever field the cursor is in.

I already have dictation built into the OS and I do not use it, because it drops the ends of sentences and I spend longer fixing it than I saved. So my questions are the boring ones:

  1. Is a locally-run model actually more accurate than the built-in dictation, or just more private?
  2. What is the latency like when the model runs on your own machine — is there a wait after you stop talking?
  3. Does any of them handle the thing that actually breaks dictation for me, which is technical vocabulary and code identifiers?

I do not need it to be free. I need it to be right often enough that I stop reaching for the keyboard to fix it.

4 answers Share
Report

Answering anonymously — a moderator will review it first.

  • @trial_week_tam · yesterday · 3 replies

    Separate the two claims, because they are not the same product decision and only one of them is about quality.

    "On device" is a privacy and offline property. It tells you the audio does not leave the machine. It says nothing at all about accuracy — a small local model can be worse than a large hosted one, and often is.

    Accuracy comes from which model, at what size, and that is the number worth asking about. Tools built on a Whisper-family model at a decent size are noticeably better at continuous speech than the OS dictation you have already rejected, mostly because they transcribe a whole utterance with context rather than committing word by word. That is also exactly why the ends of your sentences survive.

    So the honest answer to question 1 is: yes, usually better, but for a reason that has nothing to do with being local.

    26
    Share
    Reply

    Answering anonymously — a moderator will review it first.

    Report
    • @dictation_dov · yesterday

      That reframing is useful, thank you. So the question I should be asking each of these is "which model and what size", and "on device" is just the deployment.

      Does that mean the ones built on the OS speech APIs are effectively the same engine I already gave up on, wrapped in a nicer hotkey?

      12
      Share
      Reply

      Answering anonymously — a moderator will review it first.

      Report
    • @trial_week_tam · yesterday

      Roughly, yes — and that is not automatically bad. The system engine has improved a lot in recent releases, and a large part of what made built-in dictation miserable was never the recognition. It was the interaction: a dictation mode you enter, a timeout that ends it, and no clean way to put the result where you were typing.

      A hotkey that captures until you release, then pastes into the focused field, fixes the workflow without touching the model. For some people that is the entire problem solved. For you, with the vocabulary issue, probably not — that one is the model.

      17
      Share
      Reply

      Answering anonymously — a moderator will review it first.

      Report
  • @pty_paul · yesterday

    On latency, since nobody markets it: yes, there is a wait, and it is the thing that decides whether you keep using these.

    Local transcription generally runs after you stop speaking rather than while you speak, because the model wants the whole utterance. So the shape is: talk for twelve seconds, release, wait, text appears. On recent Apple silicon with a small-to-mid model that pause is short enough to live with. On an older Intel machine it is long enough that you will stop bothering.

    The practical consequence is that these suit dictating a paragraph and are wrong for dictating a word into a search box. If your imagined use is short bursts, test that specifically before paying — it is the case that feels worst and the demos never show it.

    21
    Share
    Reply

    Answering anonymously — a moderator will review it first.

    Report