A better dictation app - until Apple Intelligence catches up.
I let my WisprFlow subscription expire, to get motivated to check for alternatives. Hoping for a privacy upgrade over Flow, ideally with a known API provider I trust or local models to run.
As is often the case with AI projects, there are more startups than I care to count. What’s worse, most are just apps wrapping some third-party models. It’s hard to tell them apart. If you think about the value such apps bring, they’re in a tough spot.
Dictation should be fast, simple, accurate, and privacy-oriented. Ideally, the payment model should connect with the value the tool provides compared to the baseline. macOS’s built-in dictation is already quite good—it scores well on every point except accuracy, and even there it’s not terrible. The bar is high.
So, the question is: can you make speech transcription much more accurate without owning the model?
Flow tries to do just that, reading everything it can off your screen and sending that as context to the ASR model (and maybe an LLM afterwards). I haven’t found which model it uses. Right now it’s more accurate than Whisper Large Turbo, but you sacrifice privacy and pay a steep monthly fee.
On the other hand, there’s MacWhisper. It runs locally, has a one-time payment, and in theory checks all the boxes. But the free version’s accuracy is non-existent, and the one-time payment model feels like an app destined for future cashflow problems.
So, it shouldn’t be a monthly fee, nor a one-time payment. Nothing quite matches, except… open source. About a year ago I built a small Python app that runs Whisper locally and pastes transcriptions. But today there’s a much better alternative: VoiceInk.
An open-source app with a one-time fee (think of it as a sponsorship fee dressed up as a license):
- It already has nearly every feature of Flow.
- You know exactly where your data is going.
- And when license sales aren’t enough, the open-source community and sponsorships can keep the project alive.
Before finding VoiceInk, I stumbled on Aqua Voice, which recently released their own model: Avalon. It’s accessible via an OpenAI-compatible API and beats most benchmarks. The blog post was inspiring, so I ran a small smoke test.
Unfortunately, like most models, the results were inconclusive. A full report generated by AmpCode is here.
Funny enough, I used some examples from Aqua Voice’s blog where Avalon was supposed to shine, like: “Make a fully featured PyTorch alternative. Make no mistake.” It failed like most others, mistyping “PyTorch” as “Python.” Only ElevenLabs got it right.
Initially, I was impressed with Nvidia’s Parakeet. It runs incredibly fast on my Mac—until it started mixing other languages into English.
I ended up using Mistral through an API. It’s fast and accurate. In fact, it’s one of the most accurate models I’ve tried. Only ElevenLabs felt slightly better, but Mistral uses per-token pricing instead of a monthly fee.
At this point, VoiceInk feels like the only sensible option—until each OS eventually gets its own super-accurate dictation.