Open-Source Voice Dictation vs Wispr Flow and Friends

Open-source voice dictation has quietly become good enough that the interesting question is no longer whether a free stack can transcribe your speech accurately — it can — but who does the wiring. The Whisper family is MIT-licensed and free. Finished local apps exist and are actively maintained. Yet the paid leaders keep growing, which means they are selling something the engines do not.

That something is rarely raw accuracy. It is the hotkey, the text injection into whichever app has focus, the filler-word cleanup, the punctuation, the model download manager, the crash recovery and the support address. Those are product problems, not model problems, and they are exactly where a self-hosted stack costs you hours instead of dollars.

This guide scores the open side honestly — openai/whisper, whisper.cpp, faster-whisper, Handy, VoiceInk and FreeFlow — against Wispr Flow and superwhisper on accuracy evidence, speed, formatting intelligence, privacy posture, setup time and ongoing cost. Every price is labelled by the surface it was published on, every accuracy figure is labelled with the benchmark and date it came from, and where a widely circulated number has no disclosed methodology, we say so instead of repeating it. For the paid field on its own terms, see the companion piece, the paid-tools field guide, which puts Wispr Flow, superwhisper and Aqua Voice head to head.

Almost every disappointing experiment with open-source dictation starts with the same category error: downloading an inference engine and expecting it to behave like a product. The Whisper family is a set of libraries and command-line tools. They turn an audio file into text. They do not listen for a global hotkey, they do not know which application has keyboard focus, and they do not paste anything anywhere.

The engine layer is genuinely excellent and genuinely free. The reference implementation from OpenAI carries roughly 106,800 GitHub stars under an MIT licence at the time of writing, and the two reimplementations most people actually deploy sit beneath it. What the layer does not contain is a dictation experience — that has to come from an app on top, which you either install or write.

Star counts are a crude proxy, but the shape here is informative. The reference repository is the most starred by a wide margin because it is what everyone reads about; whisper.cpp is what a great many people actually run. If your goal is batch work rather than live typing — meeting recordings, podcast archives, call libraries — the engine layer on its own is the right answer, and we covered that build in detail in our guide to self-hosted transcription pipelines. This post is about the other job: speaking into the application you already have open.

Word error rate is the only accuracy number in this space with a shared definition, and it is published for models, not for products. That distinction matters more than it sounds: you can find a WER for Whisper Large V3 or for Cohere Transcribe, but no vendor publishes an end-to-end WER for Wispr Flow or for superwhisper as shipped products. Any single percentage attached to a dictation app should be treated as marketing until a methodology appears next to it.

A July 2026 roundup of the Hugging Face Open ASR Leaderboard puts the current open ceiling in context. Cohere Transcribe, a two-billion-parameter Apache-2.0 model, topped the board at 5.42% average WER in March 2026, and IBM Granite Speech 4.1 — also two billion parameters and also Apache 2.0 — reached 5.33% roughly five weeks later. A Whisper baseline in the same roundup sits at 7.44%.

Read that gap carefully before you draw a conclusion from it. The distance from the Whisper baseline to the best open result is 2.11 points of WER. That is real — it shows up on proper nouns, product names, acronyms and accented speech, which is precisely the vocabulary knowledge workers dictate most. It is also not the chasm the framing sometimes implies, and it comes with a licence trade-off: the newer leaders are Apache 2.0 or CC-BY-4.0, while the Whisper family is MIT with a runtime ecosystem nothing else can match.

NVIDIA Parakeet is the interesting middle. The current TDT 0.6B v3 checkpoint posts 6.32% average WER — 1.12 points better than the Whisper baseline and within a single point of the leaderboard best — at a reported real-time factor of 3,332x, among the fastest throughput figures on the board. An earlier 1.1B Parakeet variant was benchmarked at roughly 8.0% WER in a January 2026 comparison, so the line improved by about 1.68 WER points while shedding half its parameters. That is why Handy and superwhisper both offer a Parakeet path alongside Whisper.

One number deserves a warning rather than a citation. Cohere Transcribe appears at 5.42% on the March leaderboard reading and at 5.35% in superwhisper’s own July 28 post about bundling it. Those are almost certainly different model versions or different benchmark subsets rather than a contradiction, and the honest move is to carry both with their dates rather than quietly picking the flattering one.

Most open-source dictation roundups list projects and stop. The two columns that decide whether a project is usable are almost never filled in: whether it runs fully locally by default, and how much wiring you have to do before it types a word. Here is the field with both columns answered.

The column that changes decisions is the fourth one. Three of these six are engines, which means the honest cost of adopting them is not zero — it is an afternoon of wiring plus whatever maintenance the glue script needs when the operating system changes its accessibility permissions. Two are finished apps that run entirely on your machine. One is a finished app that is open source and free and still sends your audio to somebody else’s servers unless you reconfigure it.

One clarification worth making because keyword lists keep dragging it into comparisons like this one: Willow is an ESP32-based, Apache-2.0, self-hosted wake-word voice assistant — an open alternative to a smart speaker, with its own inference server — rather than a desktop dictation tool, so it is not a Wispr Flow alternative in any useful sense.

The star spread is also worth reading as a maintenance signal rather than a popularity contest. Handy sits at roughly 28,900 stars, around twelve times FreeFlow’s count, and the long tail below them contains Wispr Flow clones with double-digit or even zero star counts. Finding an open-source alternative on GitHub and finding a maintained one are different achievements, and the second is the one that matters when a macOS update breaks text injection.

The single most common error in this category is treating source-available and private-by-default as the same property. They are independent. A closed binary can run entirely on your machine. An MIT-licensed client can stream every syllable you speak to a third-party inference API. If privacy is the reason you are looking at open source at all, the licence tells you almost nothing — you have to read what the default configuration does.

FreeFlow is the clean illustration. It is free, it is open source, it is explicitly positioned as an alternative to the paid dictation apps, and its own documentation is candid about the architecture: there is no FreeFlow server, so the project stores and retains nothing, and the only information leaving your computer is the API calls to your configured transcription and language-model provider. That is an honest description of a client that, out of the box, asks for a free Groq API key and sends audio to it. The OpenAI-compatible provider setting means you can point it at a self-hosted endpoint, but that is a deliberate reconfiguration, not the default path.

Handy is the counter-example, and it is the reason it earns the free-daily-driver slot in the table above. It is built with Tauri — a Rust backend with a React and TypeScript front end — and it runs transcription through either a Whisper-family GGML/GGUF path or a CPU-optimised Parakeet V3 path, with Silero voice activity detection filtering silence before anything is transcribed. All of that happens on the machine. Distribution is through GitHub releases, a community-maintained Homebrew cask and winget, and there is a community Raycast integration on top.

The project is unusually clear about what it is optimising for. Its README states plainly that Handy is not trying to be the best speech-to-text app — it is trying to be the most forkable one — and frames the motivation as accessibility tooling belonging in everyone’s hands rather than behind a paywall. That is a genuinely different product goal from Wispr Flow’s, and it explains both the strengths and the gaps you should expect.

Three pricing models are competing here, and conflating them is how most comparisons go wrong. Wispr Flow sells a per-seat subscription. superwhisper sells a subscription with a lifetime escape hatch. VoiceInk sells a one-time licence for a binary whose source is already public. All figures below come from each vendor’s own pricing surface at the time of writing.

Run those numbers over a realistic horizon and the spread is larger than the monthly figures suggest. Three years of Wispr Flow Pro on the annual rate is $432 per seat; three years of monthly billing is $540. superwhisper’s $249.99 lifetime licence overtakes its own $8.49 monthly rate at about 29 and a half months — call it two and a half years — and costs nothing after that. VoiceInk’s $49 Personal tier is $383 less than three years of Wispr Flow Pro on the annual rate, for one Mac fewer than most teams need and macOS only.

None of which settles anything by itself, because the free tiers are more generous than the headline prices imply. Wispr Flow’s free tier allows 2,000 words a week on desktop and 1,000 a week on iPhone, with unlimited Android use described on the pricing page as a limited-time promotion. For occasional dictation that is a real product, not a trial. For heavy use it is not: one independent reviewer logged 182,718 dictated words during an extended trial, which is around 91 weeks of the free desktop allowance compressed into a single evaluation.

On speed, be careful which claim you are quoting. Wispr Flow markets dictation at roughly 220 words per minute against roughly 45 words per minute for average keyboard typing, framed as four times faster — that is a vendor comparison, not an independent benchmark. The one methodology-disclosed test we found is a single reviewer’s extended self-measurement across 182,718 words in 36 different macOS applications, which lands at a little over twice his own typing speed. Both numbers are useful; only one of them tells you what happened in someone’s real workflow.

Platform coverage is the other thing the subscription buys. Wispr Flow runs on Mac, Windows, iOS and Android for dictation, with its Notetaker meeting-transcription feature limited to Mac; it supports 100+ languages with automatic detection, and its MCP integration feeds dictated text directly into Claude, ChatGPT and other AI tools. The same reviewer notes that Linux remains the one platform gap. superwhisper covers macOS on both Apple Silicon and Intel, plus Windows and iOS. VoiceInk is Apple Silicon only. Handy is the only finished app in this comparison that covers Windows, macOS and Linux for free.

The clean version of this comparison — open source is free and private, paid software is polished and cloud-based — stopped describing the market on July 28, 2026. That is the day superwhisper announced Cohere Transcribe as a bundled local model option: an Apache-2.0 licensed, two-billion-parameter open speech model that had ranked at the top of the Hugging Face Open ASR Leaderboard. superwhisper’s post reports 5.35% average WER for it, dropping to 1.25% on clean audio, across 14 languages including English, French, Spanish, Mandarin, Japanese and Arabic. The app wraps that model with its own vocabulary and jargon handling. Both companies are based in Toronto.

Read that as a structural signal rather than a product update. The model layer in speech recognition has commoditised to the point where a paid application’s best local option is a model anyone can download for free, and the vendor’s contribution is the wrapper around it — the vocabulary handling, the model management, the app integrations. That is the same value proposition Handy declines to charge for and VoiceInk charges once for. The differentiator has moved off the model entirely.

Projecting forward, two things follow. First, per-seat dictation subscriptions will find it progressively harder to justify a model-quality story, which pushes them toward the features an open project structurally cannot match at speed: enforced enterprise controls, mobile parity, meeting capture, and integration surfaces like MCP that plug dictation into agent workflows. Second, expect more one-time licences. VoiceInk already demonstrates that a source-available app can charge for the compiled binary and the support relationship rather than for the software itself, and superwhisper’s lifetime tier is a hedge in the same direction. The subscription is the part of this market with the least defensible floor.

This is the load-bearing table. Every cell is either sourced from a vendor page, a project README or a dated benchmark, or it says plainly that no published figure exists — which is itself the most useful answer in three of the accuracy cells.

Three patterns fall out of the grid. The accuracy row is the least differentiated — four of the five options are running models from the same small pool, and the one place a real number exists for a product is superwhisper’s bundled open model. The formatting row is the most differentiated, and it is where the subscription earns its keep. The setup row is where the free options collect their real price, denominated in your afternoons rather than your card.

Notice what the empty accuracy cells imply. If no vendor publishes an end-to-end word error rate, then nobody comparing these products — including us — can rank them on accuracy with evidence. Anyone who does is either quoting a model benchmark as though it were a product benchmark, or repeating a number whose methodology was never disclosed. That is a reasonable thing to be annoyed about, and a very good reason to run a fortnight of your own dictation through two candidates before committing a team. Structuring that kind of evidence-first tool evaluation is a standard first step in our AI and digital transformation engagements.

The right answer depends far more on your constraints than on any model score. Four profiles cover most people asking this question.

For teams, the arithmetic is worth doing before the debate. Ten seats of Wispr Flow Pro on the annual rate is $1,440 a year; ten VoiceInk Personal licences is $490 once, and ten Handy installs is nothing. The subscription still wins whenever a single hour of engineering time a month would be spent maintaining the alternative — which is most of the time, in most organisations. It stops winning when the open option is already a finished app, which is exactly what has changed.

The same reasoning applies to the output side of the voice pipeline, where the open models have moved just as fast; if you are building anything that speaks back as well as listens, our guide to open-source text-to-speech covers that half. And for the broader picture of which open projects are worth a team’s attention this year, see the open-source AI project map. When the answer is to build rather than buy, our custom application work is where that wiring usually lands.