javascript-today

Voice-to-Text in 2026

OpenWhispr, Turbo Whisper, and the Right Tool for Your Platform

If you’ve ever caught yourself typing out a long technical explanation when you could have just talked through it, voice-to-text is worth a serious second look. The tooling has quietly gotten very good — and more importantly, it now runs locally without phoning home to a cloud API. Here’s a practical breakdown of the main players and which one makes the most sense depending on your OS.

The Model Underneath It All: Whisper and large-v3-turbo

Before talking about apps, it helps to understand what most of these tools are actually running under the hood.

OpenAI’s Whisper is the open-source speech recognition model that kicked off the current wave of local STT tooling. It’s trained on 680,000 hours of multilingual audio, supports 99+ languages, and is genuinely accurate in ways that older local models were not.

The model family has grown over time, and the most practically useful variant right now is whisper-large-v3-turbo. Here’s the short version of why it matters:

  • The original large-v3 has 32 decoder layers — powerful, but slow on anything that isn’t a beefy GPU
  • turbo prunes that down to just 4 decoder layers, inspired by the Distil-Whisper research
  • The result: transcription speed that’s dramatically faster with only a minor accuracy hit
  • Word error rate sits around 12% across languages — competitive with the full model in most real-world use
  • One tradeoff worth noting: turbo doesn’t support translation. If you need to transcribe non-English audio and translate it to English in one pass, stick with the standard multilingual models

For day-to-day dictation and note-taking, turbo is the right default. It’s fast enough to feel responsive on commodity hardware and accurate enough that you’re not spending time correcting transcripts.

OpenWhispr: The Polished Frontrunner

OpenWhispr has become the go-to app for engineers who want a plug-and-play dictation experience without cloud lock-in. It runs on macOS, Windows, and Linux, and the core workflow is exactly what you want: press a hotkey, speak, words appear at your cursor in whatever app is focused.

Under the hood it supports both OpenAI Whisper and NVIDIA’s Parakeet model, and gives you the option to run entirely local or wire up your own API key (BYOK) for cloud processing when you want faster turnaround on longer content.

What it does well:

  • Truly cross-platform — same experience across all three OSes
  • Fully offline mode with no data leaving your device
  • System-wide — works in terminals, editors, browsers, Slack, everywhere
  • Active development with a pipeline redesign shipping in 2026 and a custom dictionary feature for domain-specific vocabulary (useful for technical jargon, medical terms, etc.)
  • Open source, so you can audit what it’s doing

Honest caveats:

  • NVIDIA Parakeet gives you the fastest local inference, but if you’re on AMD or CPU-only hardware you’re falling back to Whisper, which is slower
  • The UI was in redesign as of early 2026 — functional but not especially refined
  • It’s still a relatively young project, so expect rough edges in edge-case workflows

If you want one tool that works across all your machines without having to think about it, OpenWhispr is the current best answer.

Going Deeper: whisper.cpp and faster-whisper

For engineers who want more control — or who are building something on top of a transcription backend — two lower-level libraries are worth knowing about.

whisper.cpp

whisper.cpp is a C++ port of Whisper from the ggml project (the same folks behind llama.cpp). It’s the most portable option:

  • Runs on CPU, Metal (Apple Silicon), CUDA, and Vulkan
  • No Python dependency
  • Outputs directly to SRT, VTT, JSON, CSV, and plain text
  • Version 1.8.3 delivered a reported 12x performance boost for systems with integrated Intel and AMD graphics

This is the right tool when you want a fast, embeddable binary with no runtime baggage. It’s also the only option that gets full GPU acceleration on Apple Silicon via Metal.

faster-whisper

faster-whisper reimplements Whisper on top of CTranslate2, a fast inference engine for transformer models. The headline number is roughly 4x faster than the original Python openai/whisper package at equivalent accuracy, with lower memory usage.

  • Python-native, integrates naturally into existing pipelines
  • Returns a generator of timestamped segments — good for building on top of
  • Supports CPU and CUDA (no Metal, so Mac performance is limited compared to whisper.cpp)
  • Best option on Linux/Windows with an NVIDIA GPU

If you’re scripting batch transcription jobs, building a custom voice pipeline, or integrating STT into another tool, faster-whisper is the practical workhorse.

Other Tools Worth Knowing

Vosk — A lightweight offline STT toolkit that predates the Whisper era but remains useful when you need broad platform support (including mobile and embedded) and small model footprints. Supports 20+ languages. Accuracy is behind Whisper but it’s fast and genuinely tiny.

VoiceInk (macOS only) — A native Swift app running from the menu bar with a global keyboard shortcut. If you live on a Mac and want the most native-feeling experience, it’s worth a look. Privacy-focused, local processing.

MacWhisper (macOS only) — More focused on transcribing audio files rather than live dictation. Good if your workflow involves dropping audio recordings and getting clean transcripts out. Polished, but not the right fit for real-time dictation.

Nerd Dictation (Linux) — A lightweight scriptable dictation tool for Linux desktops. Uses Vosk under the hood. Low resource overhead, easy to customize via shell scripts, and plays nicely with tiling window managers and keyboard-centric workflows.

Platform-by-Platform Recommendation

macOS

Best all-around: OpenWhispr
Best performance / lower-level: whisper.cpp with Metal

Apple Silicon is genuinely the best hardware to run local Whisper on right now. The Metal backend in whisper.cpp will push transcription to around 7x real-time on an M-series chip — 60 minutes of audio in roughly 8-9 minutes. For live dictation, OpenWhispr handles the UX layer well and benefits from the same acceleration.

If you’re on an Intel Mac, manage expectations — Metal is still available but slower, and you may find cloud-backed mode in OpenWhispr more responsive for real-time use.

Windows

Best all-around: OpenWhispr
Best performance / scripting: faster-whisper (with NVIDIA GPU)

Windows support in OpenWhispr works well and is the path of least resistance. If you have an NVIDIA GPU and want to build something custom or run batch jobs, faster-whisper with CUDA is the fastest option available. AMD GPU users are stuck with CPU inference for now — faster-whisper doesn’t support ROCm cleanly, and whisper.cpp’s Vulkan backend is improving but not at parity with CUDA yet.

Windows users should also keep an eye on the Vulkan path in whisper.cpp — it’s the most plausible route to proper AMD GPU acceleration.

Linux

Best all-around: OpenWhispr
Best for custom/scripted workflows: faster-whisper or whisper.cpp
Best for lightweight / keyboard-centric setups: Nerd Dictation

Linux has the most options and the most flexibility. OpenWhispr works well if you want the same experience across machines. For anything more custom — integrating STT into a workflow, running from a script, binding to a keybind in your WM — whisper.cpp or faster-whisper give you exactly the control you want. Nerd Dictation is worth knowing about if you’re on a low-resource machine or just want something minimal.

A Note on Model Size vs. Practicality

One thing that doesn’t get said enough: you probably don’t need large-v3. For most dictation use cases, whisper-small or whisper-medium with turbo gives a better speed-accuracy tradeoff than reaching for the biggest model. The small model runs in around 2 GB of RAM, the medium in ~5 GB — both are perfectly usable on everyday hardware. Save large-v3-turbo for cases where accuracy genuinely matters (technical dictation, accents, noisy environments) and you have the hardware to back it up.

Bottom Line

The local voice-to-text ecosystem has matured to the point where there’s no good reason to be routing your audio through a cloud API for everyday dictation. OpenWhispr is the practical default for anyone who wants a working setup fast. whisper.cpp and faster-whisper are the right tools once you need more control. And whisper-large-v3-turbo is the model to reach for when you need quality without waiting — just know its limitations around translation.

Pick the layer that matches what you actually need, and you’ll have a setup that’s private, fast, and works offline.