CLASSEVE
RouteHome
Best of · facts checked 2026-09-05

The best offline speech to text for Windows

Lven Instant is the best offline speech to text for Windows: on the PC, offline after setup, no word cap, never rewrites, and the only app here with a published measurement — about six times faster than you speak at a 3.76% word error rate. It is free. The ranking below covers one-time purchases, open source and voice commands.

Scored on
  1. Transcription runs on the PC, not on a server
  2. Works with the network off after setup
  3. What it does to your words
  4. Usage limits and price shape
  5. Published speed and accuracy
The ranking
  1. Lven Instant

    ClassEve

    The only dictation product compared on this site that is always on-device, never rewrites your words, has no usage cap and ships on Linux — and it is free.

    Right for: Daily dictation on a Windows PC, offline, with the same account on Linux and Android.

    Where it transcribes
    On your device — nothing uploaded
    Your words
    Typed as said, punctuated, never rewritten
    Speed and accuracy
    About 6× faster than you speak (RTF 0.166), 3.76% word error rate on clean speech — consumer-grade hardware, 73 clips, measured 2026-07-23.
    Limits
    None
    Platforms
    Windows, Linux, Android
    Price
    Free
  2. On-device push-to-talk dictation for Windows only, as a $29 one-time purchase.

    Right for: One Windows machine and the lowest one-time price.

    Where it transcribes
    On-device (per its site)
    Platforms
    Windows only
    Price
    One-time purchase ($29, per its site)
    Speed
    No published measurement
  3. On-device dictation for Mac and Windows as a one-time purchase.

    Right for: One person, one Mac or Windows machine, a single payment.

    Where it transcribes
    On-device (per its site)
    Platforms
    macOS and Windows. No Linux, no Android.
    Price
    One-time purchase ($79, per its site)
    Speed
    No published measurement
  4. Free, MIT-licensed Whisper dictation that runs offline, for people happy to manage models themselves. On the same hardware where Lven Instant's engine runs about six times faster than you speak, whisper-large-v3 never reaches realtime.

    Right for: Free and open source above everything, and comfortable troubleshooting models and configs yourself.

    Where it transcribes
    On-device (Whisper)
    Platforms
    Windows, macOS, Linux (per its site). No Android.
    Price
    Free, MIT
    Support
    Community and repository issues
  5. The professional suite: deep voice commands, corrections by voice, medical and legal vocabularies, at professional prices and with a real learning curve.

    Right for: Hands-free control of the whole computer, or a workplace that already standardizes on Dragon.

    Where it transcribes
    Primarily local on desktop editions; varies by product line
    Platforms
    Windows desktop; separate mobile products
    Price
    Professional pricing — typically hundreds of dollars per license
    Speed
    No published measurement
  6. Free, built in, and on most configurations it sends your audio to Microsoft's online speech service.

    Right for: Occasional dictation with zero setup and zero cost, when cloud processing is not a concern.

    Where it transcribes
    Online speech services on most configurations
    Works offline
    Generally requires a connection
    Platforms
    Windows only
    Price
    Free
FAQ

Common questions.

Does Windows have built-in offline speech to text?
Windows voice typing (Win+H) relies on Microsoft's online speech service on most configurations and generally needs a connection; on-device support varies by hardware generation. For dictation that is offline by design, use an app that runs its own model on the PC, such as Lven Instant.
Does offline speech to text need a GPU?
Not Lven Instant. Its engine runs on the CPU in about 600 MB of memory, and the published measurement was taken on consumer-grade hardware. whisper-large-v3 on the same hardware needed a GPU with 3.36 GB of video memory and still ran slower than realtime.
Is anything sent to a server at all?
With Lven Instant, sign-in, the one-time model download and access checks use the network. None of those requests carry audio or transcript text.