Picking a transcription app shouldn't feel like pulling teeth.

Buzz is one of those tools that solves exactly one problem and then gets out of your way. The problem in question — running OpenAI's Whisper model on a file you don't want leaving your laptop — used to involve at least one virtual environment, a couple of pip install commands, and a 50-50 chance of an opaque ffmpeg error. Buzz turns all of that into a window with a button labeled "Transcribe". Drag in your audio, pick a model size, watch the progress bar. Output goes to a folder. Done.

That's the whole pitch. If it sounds underwhelming, that's because the surface is supposed to be underwhelming. The interesting bit is what it doesn't do — it doesn't ask you to make an account, doesn't push your file to a server, doesn't try to upsell you to an "AI summary" feature on a credits plan. It is genuinely just a wrapper. And after spending a fair bit of time with the app across two laptops and three different operating systems, I think that minimalism is the strongest thing about it.

What Buzz actually is

Buzz is an open-source desktop application written by Chidi Williams. The repository on GitHub has been around since 2022 and continues to receive updates. Under the hood, it bundles two transcription engines: the original OpenAI Whisper implementation in Python, and the much faster whisper.cpp port written in C++. You can pick which one you want at the model-loading screen — and which one you want depends on what kind of computer you're sitting at.

If you've ever used a tool like Audacity, Buzz will look familiar in spirit: utilitarian, slightly old-fashioned in its widgets, clearly designed by people who care more about the verb than the noun. There's no marketing-driven empty space. There's no dashboard. The main window is a list of transcription jobs, and each row tells you whether it's queued, running, or finished.

How it feels to use

The first thing that struck me was how quickly I got from "I just downloaded this" to "I have a transcript." On a 2021 M1 MacBook Pro, the entire setup took about three minutes — and most of that was the initial model download (the medium-size model is about 1.5 GB). On a five-year-old Windows laptop without a discrete GPU, it took longer to run the transcription, but the setup was identical.

The interface isn't beautiful. I want to get that out of the way. It's built with PyQt, and it looks like every PyQt app ever made: serviceable, slightly chunky, with menus that haven't been redesigned since the early 2010s. But I've spent enough time with truly bad audio software to know the difference between "ugly" and "clumsy", and Buzz is firmly in the first category. Things are where you'd expect. The keyboard shortcuts work. The progress reporting is honest — the percentage moves at roughly the speed it claims to be moving.

I tested it on a 47-minute interview I'd recorded for an unrelated project. Tiny model finished in about 90 seconds and produced a transcript that was readable but not really usable; about one word in twenty was wrong. The base model took maybe four minutes and the result was significantly better — I'd say good enough for a personal note-taking workflow. The medium model took around fifteen minutes on the M1, and the output was the kind of transcript you'd want if you were going to publish it after a light edit. None of this should surprise anyone who's used Whisper before. The point is just that Buzz exposes those tradeoffs in plain English instead of asking you to memorize model names.

The best thing I can say about Buzz is that I forgot I was using it. It does the job, the job ends, the file appears.

Under the hood

Buzz lets you choose between several model sizes (tiny, base, small, medium, large) and several backends. The most important choice is between the original OpenAI Whisper Python implementation, the whisper.cpp backend, and Hugging Face's transformers-based implementation. There's also support for using OpenAI's hosted Whisper API if you'd prefer to send the file to OpenAI in exchange for faster results — but that defeats the privacy advantage, and almost no one I know who installs Buzz uses that mode.

Two practical notes from real-world use:

• On Apple Silicon, the whisper.cpp backend with Core ML acceleration is the fastest by a wide margin. You'll want to enable that.
• On any machine with a recent NVIDIA GPU and CUDA installed, the original PyTorch backend will use the GPU and become noticeably faster than the CPU-only paths. On laptops without a dedicated GPU, the difference between backends shrinks dramatically.

Buzz also supports a "Live Recording" mode where it'll transcribe directly from your microphone as you speak. I've used this feature exactly twice, and both times I came away thinking that this is not what Buzz is for. The latency is wrong for it — you'll get text in chunks of several seconds — and it doesn't integrate with other apps. If you want dictation that drops text where your cursor is, look at VoiceInk instead. If you want live captions for a video call, look elsewhere entirely. Buzz is a file-based tool with a microphone option grafted on, and you can feel the seam.

Pros and cons

How to use Buzz: a step-by-step

This is the part most write-ups skip, so here it is. The whole flow, from "I haven't installed anything" to "I have a clean SRT", on a Mac. Windows and Linux are nearly identical.

1. Download Buzz from the official GitHub releases page. Pick the build for your operating system. On macOS, that's a .dmg; on Windows it's an .exe installer; on Linux you've got AppImage and Snap options.
2. Install it like any normal app. On Mac, drag to Applications. On first launch, macOS may complain about an unidentified developer; control-click and choose "Open" once and the warning goes away.
3. Open Preferences and choose your default backend. On Apple Silicon, "Whisper.cpp" with Core ML support is the right answer. On Windows with an NVIDIA GPU, the OpenAI Whisper backend will use CUDA. Otherwise, leave it on the default.
4. Drag your audio or video file into the main window. Buzz accepts MP3, WAV, M4A, FLAC, MP4, MOV — basically anything ffmpeg can read.
5. Pick a model size. Start with base if you're not sure. Move up to medium for cleaner output. The large model is the most accurate but slow and memory-hungry.
6. Pick output formats. For interviews, TXT plus SRT is the right combination. The first is for reading; the second is for any future cleanup work in a subtitle editor.
7. Hit Transcribe and walk away. Seriously. Make a cup of tea. The progress bar updates honestly. When it finishes, the output files appear next to the source.

Compared to the alternatives

If we're staying inside this site's shortlist, here's where Buzz sits relative to the others:

Versus Subtitle Edit: Subtitle Edit can also drive Whisper, but it does much more than transcribe — it has a full waveform editor and supports an absurd number of subtitle formats. If you're a translator or a captioner, Subtitle Edit is probably your daily driver and Buzz is redundant. If you just want a transcript, Buzz is faster to learn.

Versus Whisper Transcription (Mac): Whisper Transcription is more polished, prettier, and better integrated into macOS. It's also Mac-only and has a paid tier. Buzz is uglier, but free everywhere.

Versus Pyrenees: Pyrenees is faster on Apple Silicon, full stop — but only on Apple Silicon. If you're on an M-series Mac and you mostly transcribe shorter files, Pyrenees wins on speed. Buzz wins on cross-platform consistency and on having more backend options.

Versus VoiceInk: Different tool for a different job. VoiceInk is for live dictation (talking into apps as you'd talk into iOS dictation). Buzz is for files. They don't really compete.

Who should use Buzz

FAQ

I want to start with a confession. The first time I tried to use Subtitle Edit for transcription, I gave up after fifteen minutes and went back to Buzz. The interface looked like a Windows XP control panel that had been dragged forward by sheer force of will, and I couldn't figure out how to make Whisper run from inside it. I assumed it was buggy. It wasn't. I was just wrong about what kind of program I was looking at.

Subtitle Edit is not a transcription app with caption editing tacked on. It's a caption editor with transcription tacked on. The difference matters. The whole layout — the waveform on the bottom, the cue list on the right, the keyboard-shortcut-everything ergonomics — assumes you already have something resembling subtitles and you're working on them. Once that flipped in my head, the program suddenly made sense, and within an hour I was using it for things I hadn't realized I needed.

Two ways to think about it

There are essentially two distinct workflows the app supports, and people who try Subtitle Edit usually fall into one camp or the other.

Workflow A: you have an audio or video file and you want clean, correctly-timed captions. You import the media, you tell Subtitle Edit to run Whisper on it, you wait. You get a populated cue list. Then you spend twenty or thirty minutes cleaning it up: fixing punctuation, merging short cues, splitting long ones, retiming the parts where the model got confused. The output goes out as .srt or whatever else you need. This is the workflow professional captioners use.

Workflow B: you have a transcript already (from Buzz, from Whisper Transcription, from Otter, doesn't matter) and you want to fix it. You open the existing file, you bring in the audio so the waveform syncs with the cues, and you fix the obvious mistakes by listening and clicking. This is what I personally use it for, and I think it's the underrated use case. Even if your primary transcriber is something else entirely, Subtitle Edit makes a phenomenal "second tool".

What's it like to actually use?

The interface is dense. There's no other word for it. Every pixel earns its keep. On first launch you see a menu bar with maybe a dozen drop-downs, a toolbar with another dozen icons, and three separate panels that all want your attention at once. If you're used to single-purpose apps with three buttons total, the first impression is overwhelming.

Stick with it for an afternoon and the density becomes a feature. The reason every command has a keyboard shortcut is that captioners burn through hundreds of cues per session and need to keep their hands on the keyboard. The reason the waveform is always visible is that you're going to be matching cue boundaries to audio events, and it's faster to see them than to scrub. The "ugly" interface is the price of being able to do real work.

Whisper integration sits under Video → Audio to text (Whisper). From there you choose the engine — Subtitle Edit supports the original Python Whisper, whisper.cpp, Const-me's GPU implementation, Purfview's Whisper Faster, and a couple of others depending on which version you have installed. Each engine has its own strengths. On a Windows laptop without a GPU, Purfview's implementation gave me the best balance of speed and accuracy. On a machine with an NVIDIA card, Const-me's GPU build was faster than anything else by a wide margin.

The format support is genuinely absurd

This is the part nobody talks about, and the thing that makes Subtitle Edit irreplaceable in some workflows. The app reads and writes well over two hundred subtitle formats. If you've ever stared at a file with a strange extension and wondered how to convert it to .srt without losing timing or styling, Subtitle Edit is almost certainly the answer.

A non-exhaustive sample of what it'll handle:

• The everyday ones: SRT, VTT (WebVTT), ASS, SSA, TXT
• Broadcast formats: EBU STL, Cavena 890, EBU-TT, EBU-TT-D, IMSC 1.1, TTML, DFXP, SCC, SMPTE-TT
• DVD/Blu-ray: VobSub (.sub/.idx), SUP (BD), and OCR support for both
• Streaming/professional: Netflix's IMSC variant, Apple iTT, MicroDVD, MPL2
• Weird and old: JSON Web, FAB, Sonic Scenarist, ATS, plus a long tail of pretty obscure formats

If your transcription job ends in "and then we hand it to a broadcaster," Subtitle Edit may be the only free tool that can produce the format the broadcaster wants. That alone is reason enough for some people to keep it installed.

Where it stumbles

Honest section. Subtitle Edit is not for everyone, and there are real friction points beyond the dense UI.

The platform story is uneven. The Windows version is a proper, signed, native application that has had two decades of polish. The Mac version is a newer port that runs natively on both Intel and Apple Silicon, but feels less mature — keyboard shortcuts that work flawlessly on Windows occasionally do nothing on Mac, certain dialogs appear off-screen, and waveform extraction sometimes fails on file types that work fine on the Windows build. On Linux, you're typically running it through Mono, which works but has its own assortment of papercuts. If you're not on Windows, expect rougher edges.

It's not a transcription-first app. If your goal is to get a clean .txt transcript and you don't care about timing, you'll find yourself fighting a UI that wants you to care about timing. You can absolutely use it for plain transcripts — just export to TXT after the cues are populated — but you'll spend a lot of attention on widgets you didn't need.

The translation features are uneven. There are translation integrations (Google, DeepL, libretranslate, ChatGPT API, others), but the quality varies and the UX of running them feels grafted on. For pure translation work, you're better off elsewhere.

The learning curve is real. Out of every tool we cover on this site, Subtitle Edit has the steepest first-week curve. Plan for it.

Pros & cons, condensed

How to use it: three short walkthroughs

Because Subtitle Edit covers more than one workflow, a single step-by-step guide doesn't quite work. Here are three focused ones.

1. From audio file to clean SRT (full Whisper workflow)

1. Open Subtitle Edit and load your video or audio. File → Open video file. The waveform should appear at the bottom; if it doesn't, the file extension may not be supported and you'll need to convert.
2. Open the Whisper dialog: Video → Audio to text (Whisper). Pick the engine and model size. Faster-Whisper with the medium model is a good default on most Windows machines.
3. Let it run. Progress shows in a small log window. A 30-minute file on a midrange Windows laptop typically completes in 8–15 minutes.
4. Review the cue list. Each cue is a row. The waveform shows where it sits in the audio. Use the keyboard arrows to step through cues; spacebar plays the current one.
5. Fix obvious problems. Common ones: cues that are too long (use Tools → Split lines longer than…), overlapping cues, wrong proper nouns. Most of this can be handled with built-in batch tools.
6. Export. File → Save as → SRT for general use, or whichever format your downstream tool needs. Subtitle Edit can also do batch convert in case you have many files.

2. Cleaning up an SRT that another tool produced

1. Open the existing SRT in Subtitle Edit (File → Open).
2. Bring in the original audio via Video → Open video file. The waveform now syncs with the existing cues.
3. Run "Fix common errors" from the Tools menu. This catches things like missing spaces after periods, capitalization issues, double spaces, and Whisper's habit of starting cues with hesitation markers.
4. Walk through the cues that look suspicious — usually the very long ones and the very short ones — and fix them by listening to the audio.
5. Save. Done.

3. Extracting subtitles from a Blu-ray rip

1. File → Import → VobSub or Blu-ray sup, depending on what you have.
2. Choose the OCR engine. The built-in nOCR is decent for European languages; Tesseract is better for some scripts.
3. Run OCR. Review the results and correct any glyphs the engine misread.
4. Save out as SRT for use anywhere else.

Subtitle Edit vs. the others

If we line it up against the rest of the shortlist:

Compared to Buzz, Subtitle Edit is the heavy tool. Buzz is for "I have a recording, I want a transcript". Subtitle Edit is for "I have a recording, I want broadcast-ready captions, and I'm willing to spend an afternoon getting them right." Both are free; they're answers to different questions.

Compared to Whisper Transcription, Subtitle Edit is dramatically uglier and dramatically more capable. Whisper Transcription will get you a clean transcript faster on a Mac. Subtitle Edit will let you actually shape it.

Compared to Pyrenees, the comparison doesn't really hold — Pyrenees is a transcription engine optimized for speed, Subtitle Edit is an editing environment. They could even live alongside each other: Pyrenees produces, Subtitle Edit edits.

Compared to VoiceInk, they share no overlap at all. Different jobs.

Who it's for

FAQ

Setting expectations

This is one of the rare apps in this space that someone clearly designed, rather than just shipped. You can tell immediately. The icon doesn't look like a Python logo with a microphone slapped on. The window has the right corner radius. The settings panel uses the macOS sheet style that lets you actually find what you're looking for. When you import a file, the app shows you metadata — sample rate, channels, length — that other transcription tools just shrug at.

None of this changes the underlying transcription quality. Whisper is Whisper, regardless of which app calls it. So the question Whisper Transcription has to answer is: given that the model is the same, what does this app give me that the free options don't?

The honest answer, after spending a couple of weeks with it: a lot of small things, none of which are individually decisive, but which add up to "this is the app I'd hand to my mother."

What it actually does

The core flow is identical to every other tool in this category. Drag a file in, pick a model, press a button, get text. Where Whisper Transcription differentiates is in the small touches.

The transcript view is interactive. Click a sentence, the audio jumps to that timestamp. Edit the sentence in place. Highlight a span and you get inline tools to merge cues, split them, change capitalization, mark a speaker. It's not Subtitle Edit's level of cue-editing power, but for working with prose-style transcripts, it's genuinely faster than re-opening your output in another app.

It can capture system audio, not just microphone. A small but uncommon feature. If you want to transcribe a YouTube video, a podcast you're listening to, or a Zoom call (with appropriate permissions), Whisper Transcription can pipe the system's audio output directly in. Most of the free alternatives only see the microphone.

Export is well thought through. SRT, VTT, plain text, and DOCX are all one click away. The DOCX export in particular is more polished than what you'll get from running Whisper through a script — it preserves paragraph breaks at sensible points, includes timestamps as headers if you want them, and doesn't dump everything into a single block of unreadable prose.

There's a menu-bar mode. If you click the menubar icon, a small palette appears that lets you start a recording, drop in a file, or pull up your recent transcripts without opening the main app. It's the kind of detail a tinkerer never builds and a designer always insists on.

The pricing question

This is the part of the review where we have to talk about money, because it's the main thing that separates Whisper Transcription from the free alternatives.

The app is free to download from the Mac App Store. The free tier comes with the smaller Whisper models — typically tiny and base — which are fine for casual notes but noticeably less accurate than what you'd want for professional work. To unlock the larger models (medium, large, and the various distilled variants), you make a one-time in-app purchase. Pricing has shifted over time and may differ by region, so I'd rather you check the App Store listing than trust a number we wrote down at one point.

What's worth saying is that the pricing model is a one-time unlock, not a subscription. You pay once and you own the larger models. There's no monthly fee, no per-minute charge, no credits system. That alone makes it cheaper than most cloud-based transcription services if you transcribe more than a few hours a month.

Where the polish ends

I want to be direct about the limits here, because every "the polished one" review I've ever read tends to gloss over them.

Mac only. Obvious but worth saying. If you ever switch to Windows or Linux, your purchase doesn't follow you and your workflow doesn't follow you.

Less flexible than open-source alternatives. The app picks reasonable defaults and hides most of the tuning knobs. If you want to set custom Whisper parameters, run a fine-tuned model, or experiment with non-standard backends, you'll outgrow Whisper Transcription quickly. Buzz lets you switch backends; this doesn't.

Speed is good but not the best. On Apple Silicon, Pyrenees is faster — sometimes substantially faster — for the same model size. Whisper Transcription uses solid acceleration but isn't the speed champion of the field.

No deep subtitle editing. The interactive editor is a pleasure for prose, but it's not pretending to be Subtitle Edit. If your job involves cue-by-cue caption work, you'll still be exporting to .srt and finishing the job elsewhere.

App Store review constraints. Because it's distributed through the App Store, it lives inside Apple's sandbox rules. That has security upsides (the app can't quietly access files you didn't grant it access to) but the occasional UX papercut — for instance, you'll be re-asked for microphone permission after some macOS updates.

Pros and cons

How to actually use it

The flow is shorter than for most tools we've reviewed. Here's the abridged version.

1. Install from the Mac App Store. Search "Whisper Transcription" and install. No external installer, no permissions juggling.
2. Open it and let it download the default model. The free models are small enough that this is fast.
3. Drop in a file or click the record button. Audio and video files work; the app strips audio automatically.
4. Pick the model and language. If you've unlocked the larger models, medium is a sweet spot for most use cases. Language can be left on auto-detect.
5. Start the transcription. Watch the progress bar — or, more usefully, switch to another app and ignore it until it's done.
6. Edit the transcript inline. Click any sentence to play it back. Fix mistakes. Tag speakers.
7. Export. File → Export, pick the format. Done.

Compared to the others

Quick reference points across the rest of the shortlist:

Versus Buzz: Buzz is free everywhere; Whisper Transcription is a paid Mac app. If you're disciplined enough to set up Buzz and don't mind its plain UI, you get the same transcription quality without spending anything. If you want it to feel like a Mac app and you transcribe regularly enough that the time savings matter, the purchase pays itself back.

Versus Pyrenees: Pyrenees is faster and free, but barer-bones. No interactive editor, no DOCX export, no system audio capture. If raw speed and zero cost are your priorities, Pyrenees. If polish is your priority, this.

Versus Subtitle Edit: Different category. Whisper Transcription is for getting transcripts; Subtitle Edit is for grooming captions. If you do both, you'll likely use both.

Versus VoiceInk: Different again. VoiceInk is for live dictation into other apps. Whisper Transcription is for files (with optional recording). They cover different problems.

FAQ

What it is, in two paragraphs

Pyrenees is a free macOS transcription app for Apple Silicon Macs. It's built around MLX, Apple's open-source machine-learning framework released in late 2023. MLX is designed specifically for the unified-memory architecture of M-series chips — it runs models on the GPU and Neural Engine without copying tensors back and forth across separate VRAM and system RAM the way frameworks designed for NVIDIA cards have to. For models like Whisper, that translates into noticeably faster inference than running the same model through plain PyTorch or even whisper.cpp's Core ML path.

The app itself is small, quiet, and does basically nothing besides transcription. You import a file, pick a model, get a transcript. There's a list of recent jobs. There's an export menu with the usual SRT, VTT, and TXT options. There is no waveform editor, no system audio capture, no menubar palette, no fine-tuning of model parameters beyond a few sensible toggles. The minimalism is deliberate.

So how fast is it, really?

Here's where I have to flag the usual caveats. Speed comparisons across transcription apps are extremely hardware-dependent, and any specific number we put on a page will be wrong by the time someone reads it. So instead of giving you "Pyrenees finished a 30-minute file in X seconds while Buzz took Y", which would be misleading without knowing your Mac, I'll describe the pattern we observed.

On every Apple Silicon Mac we tested — an M1 MacBook Air, an M2 MacBook Air, and a Mac Studio with an M2 Max — Pyrenees ran the same Whisper model meaningfully faster than Buzz did, and modestly faster than Whisper Transcription. The gap was widest on the larger models (medium and large) and smallest on the tiny model, where everything is fast enough that the choice of framework barely matters.

The qualitative experience is the more interesting part. Where transcription on the same hardware in Buzz used to feel like "start the job, go do something else for a quarter of an hour", Pyrenees turns it into something more like "wait for it to finish while you're still on the page". For shorter files — voice memos, meeting fragments, ten-minute lectures — this changes the workflow. You stop batching things and you start transcribing them ad hoc.

Pyrenees doesn't feel faster the way an upgraded computer feels faster. It feels faster the way switching from email-based reviews to instant messaging feels faster — the activation energy drops below a threshold, and you do more of it.

What's actually different vs. the others

Most Mac transcription apps fall into one of two technical camps. They either ship the original PyTorch Whisper implementation with whatever GPU acceleration they can scrape together, or they bundle whisper.cpp, the C++ port that Georgi Gerganov maintains. Both are perfectly good options.

Pyrenees is in a third camp. It uses MLX-converted Whisper weights and runs them through the MLX runtime. Because MLX is built specifically for the M-series unified memory model, it can do things like memory-map the model weights without copies, run quantized variants natively, and use the hardware in ways that more general-purpose frameworks can't quite manage.

The user-visible consequences:

• Less memory pressure. Running the large model on a 16 GB Mac is genuinely viable in Pyrenees. In some other apps with the same model, you'll see the system swap heavily and the app slow to a crawl on long files.
• The Neural Engine is actually used. A lot of "Apple Silicon optimized" apps pay lip service to the ANE without really hitting it. MLX engages it for the parts of Whisper where it helps.
• Quantized models load quickly. Pyrenees ships several quantized variants — 4-bit, 8-bit — that produce nearly the same accuracy as the full-precision models at a fraction of the resource cost.

What Pyrenees doesn't do

This is the section where the case for a different tool gets made. Pyrenees is opinionated about its scope, and the things it won't do are choices, not oversights.

It doesn't run on Intel Macs. Apple Silicon only. If you're still on a 2019 16-inch MacBook Pro, this app is not for you, and you'll have to wait for that next upgrade or fall back on Buzz.

It doesn't run on Windows or Linux. Obvious from the platform note, but worth saying if you're considering setting up a multi-OS workflow.

It doesn't have an interactive transcript editor. The output is a finished transcript. You can fix obvious typos in the export, but there's no click-to-play, no inline cue editing, no speaker labels.

It doesn't capture system audio. Microphone input works for ad-hoc recording, but it can't pull audio from another app the way Whisper Transcription can.

It doesn't do dictation. File-based only. For dictation into other apps, you want VoiceInk.

It doesn't have professional subtitle exports. SRT, VTT, and TXT cover the common cases. If you need TTML, EBU-STL, or any of the broadcast formats, you'll need to take the SRT into Subtitle Edit for conversion.

Pros & cons

How to use it

This is the shortest "how to" of any tool we cover, because the app is genuinely that simple.

1. Download Pyrenees from the developer's site (look for the official notarized DMG; avoid mirrors).
2. Drag the app to /Applications. First launch will hit Gatekeeper; right-click the icon, choose "Open", confirm. macOS will remember the choice.
3. Pick a model on first launch. The app will offer to download one. Medium 4-bit is a good first choice on most M1/M2 Macs.
4. Drop your audio or video file in. Or hit the record button to capture from the microphone.
5. Confirm the language if you don't want auto-detection.
6. Wait. Briefly. Pyrenees is faster than the alternatives; you'll often be done before you finish your coffee.
7. Export. File → Export → SRT/VTT/TXT.

How it compares to the rest of the shortlist

Versus Buzz: Pyrenees is faster and prettier. Buzz is more flexible (Linux, Windows, multiple backends, batch queueing, OpenAI API support). For Mac-only users who don't need Buzz's flexibility, Pyrenees wins. For anyone who works across platforms or needs the optionality, Buzz still has a place.

Versus Whisper Transcription: Pyrenees is faster and free; Whisper Transcription is more polished and has features (interactive editor, system audio, DOCX export) that Pyrenees doesn't. It's a real tradeoff. Try Pyrenees first since it costs nothing — if you need the extra features after a week of use, Whisper Transcription's purchase makes sense.

Versus Subtitle Edit: Different jobs. Pyrenees produces, Subtitle Edit edits. The natural workflow is to use both.

Versus VoiceInk: Different jobs again. Pyrenees is for files, VoiceInk is for live dictation.

FAQ

The thing it replaces

Apple has shipped dictation as a system feature on macOS for over a decade. You press a hotkey and you can speak into any text field. It works. It's been there for so long that most people have forgotten about it. The reason VoiceInk exists is that the built-in version has some annoying limitations — it sends your voice to Apple's servers (or used to; recent versions can run on-device for English on Apple Silicon, but the implementation is opaque), it doesn't always handle technical vocabulary well, and the customization is essentially zero. You can't tune it for your jargon, your accent, or your workflow.

VoiceInk replaces that with something more flexible. It's an open-source app that runs Whisper locally and gives you a hotkey to dictate with. The text appears wherever your cursor is. The model is on your machine. The customization is yours. It's free, the source is on GitHub, and once you've used it for a couple of days, going back to system dictation feels like an unnecessary downgrade.

The interaction, in detail

This is the only review where the actual gesture matters more than the underlying tech, so it's worth describing carefully.

You set a hotkey in VoiceInk's preferences — let's say fn, the function key, since it's already on your keyboard and it's not bound to anything most people use regularly. From then on, you can be in any application — a browser, a terminal, an email client, a code editor — and:

1. Hold the hotkey down. A small overlay appears at the bottom of the screen showing that recording has started.
2. Talk. The overlay shows a moving waveform so you know the mic is active. Talk normally; the model is forgiving of "um" and "uh" and natural speech rhythm.
3. Release the hotkey. A second or two of processing happens (faster if you have a small model loaded, slower if you have a large one), and then the transcribed text gets typed into wherever your cursor was. As if you'd written it yourself, just much faster than you can type.

That's the whole interaction. Hold, talk, release. Once it's part of your muscle memory, it changes how you write a lot of small things — Slack replies, email responses, code comments, search queries. I've watched people pick it up for the first time and stop using their keyboard for short messages within an afternoon.

Where the model lives

Like the other tools on this site, VoiceInk runs Whisper locally. It supports several model sizes, and the choice of model is a real-time-vs-quality tradeoff that matters more than it does for file-based transcription. With dictation, you don't want to wait twenty seconds for the text to appear; you want it now. So most VoiceInk users settle on either the base or small model, which transcribe quickly enough to feel responsive.

That tradeoff has a downside. Smaller Whisper models are noticeably less accurate than the large variants, especially with proper nouns, technical jargon, or non-mainstream accents. VoiceInk has a "vocabulary" feature where you can teach it specific words you use often — names of people you work with, project names, technical terms — and that helps a lot. But the accuracy gap is real, and you should expect to do light corrections after dictating anything important.

What VoiceInk is genuinely good at

I want to be specific about this rather than vague, because vague reviews are useless when you're deciding whether to install something.

Quick replies. Slack messages, email responses, "yeah looks good", "let me get back to you tomorrow on that." Things that take three seconds to say and ten seconds to type. The hotkey workflow shaves real time off a real day.

First drafts. Talking out a paragraph and then editing it on the keyboard is genuinely faster than writing the paragraph from scratch for many people. VoiceInk fits this workflow especially well because the text lands directly in your editor of choice — no copy-paste step.

Note-taking. Quick thought, capture it before it goes away. The hotkey-anywhere model means the friction of "where is my notes app, where is the cursor, what was I about to say" disappears.

Code comments and commit messages. Anywhere a thought is more important than its phrasing. The fact that you can be in your terminal or your editor makes this work without breaking flow.

Accessibility. For people with RSI, hand injuries, or other reasons keyboard input is painful, a fast on-device dictation tool is genuinely valuable. VoiceInk's open-source nature also means you can audit it for any concerns about where your voice goes.

What it isn't good at

Long-form writing. An hour of dictation is exhausting in a way an hour of typing isn't. People who try to dictate entire essays usually go back to keyboards within a week.

Anything you don't want misheard. If accuracy is critical — a legal document, an academic citation, a medical reference — dictation will let you down at small but inconvenient frequency. Always read it back.

File transcription. Said it once, saying it again. VoiceInk doesn't accept input audio files. It records from your microphone, processes it, and types the result. If you have a file, use Buzz or Pyrenees.

Multi-speaker situations. The mic captures whoever is loudest. A meeting recording is the wrong input for VoiceInk.

Pros & cons

Setting it up

1. Install VoiceInk from its GitHub releases page or the developer's site.
2. On first launch, grant permissions. macOS will ask for microphone access and accessibility access (the latter is what lets the app type into other apps). Both are required.
3. Pick a model. Small is a good starting point; bump up to medium if you have an M-series Mac and want better accuracy.
4. Set a hotkey you'll remember. The function key (fn) works well because it's not used for much. Some people prefer right-Option, which is also unused on most layouts.
5. Add custom vocabulary if you have specialized words. Names of coworkers, project codenames, technical terms. The model uses these as hints during transcription.
6. Try it in a low-stakes app first. Open a TextEdit window, hold the hotkey, dictate a paragraph, see how it goes. Adjust your speaking pace based on what happens.

Compared to the rest of the shortlist

VoiceInk is genuinely orthogonal to every other tool on this site. The other four answer "I have audio, give me text." VoiceInk answers "I have a thought, give me text faster than I can type." Those are different problems and different tools, and they're not in competition.

If you've installed VoiceInk and you find yourself wanting to transcribe a meeting recording, you're using the wrong app — go install Buzz or Pyrenees for that. Conversely, if you've installed Buzz and you find yourself thinking "I wish I could dictate into Slack the same way," go install VoiceInk. It's normal to have both.

FAQ