Wordmate
An on-device speech-to-text app for Apple Silicon Macs that turns raw microphone audio into clean, formatted text — using local AI models to strip filler words and stutters without sending any audio to the cloud.
🔗 Visit Wordmate
Description
If you've ever tried dictating a document by talking to your phone, you know the result is usually a mess — half-finished sentences, "um"s and "uh"s, words repeated twice, no punctuation anywhere. Wordmate is built to fix exactly that: talk into your Mac's microphone, and instead of getting a rough transcript you have to clean up by hand, you get text that already reads like something a person wrote and proofread. Think of it as a personal stenographer that tidies up your speech as it types — and unlike most dictation tools, which send your voice to a company's servers, everything happens directly on your own computer, so nothing you say ever leaves your machine.
Under the hood, Wordmate pairs NVIDIA's Parakeet TDT v3 speech-recognition model with a small Qwen 0.6B language model, both running locally via Apple's Metal GPU acceleration on Apple Silicon. Audio is processed in near real time (every 5 seconds during recording), with Parakeet doing raw transcription and Qwen handling cleanup — stripping filler words and false starts, converting spoken numbers to numerals, splitting long dictation into paragraphs, and applying spoken corrections. At release the pipeline reports a 79-91% "clean words" rate (text needing no further edit) with sub-300ms median latency, plus safety checks meant to catch model hallucinations before they land in the transcript. The code is open source on GitHub; the packaged app is distributed separately at wordmate.sh, and its pricing isn't disclosed in the repository.
💬 Our review
The short version: if you have an Apple Silicon Mac and want dictation that doesn't ship your voice to a server, Wordmate is a genuinely interesting, privacy-first option — though it's young, Mac-only, and its commercial pricing isn't public yet.
Most competent dictation tools today are either OS-level (macOS's built-in dictation), cloud-based services like Otter.ai (roughly $8-20/month), or the wave of Whisper-based local apps that already dominate this exact niche: Superwhisper (around $8.49/month or a one-time license option) and MacWhisper (one-time purchase, roughly $79 for the Pro tier) both run OpenAI's Whisper locally on Apple Silicon, and both have a multi-year head start with larger user bases and more polished UIs. Wordmate's actual differentiator is its two-model pipeline — Parakeet TDT v3 for transcription plus a small local Qwen model specifically tuned to strip filler words, fix stutters, and reformat spoken text — which is a more deliberate approach to "clean" output than a raw Whisper transcript. The claimed 79-91% clean-word rate and sub-300ms latency are solid numbers, but they're self-reported and not independently benchmarked. The real catch: the code is open source, but the actual product lives behind a separate commercial app (wordmate.sh) whose price the repository never states, so there's no way to make an apples-to-apples value call against Superwhisper or MacWhisper's known pricing. Worth trying if privacy is a hard requirement and you're comfortable with an early-stage, single-developer project; if you just want the most polished experience today, MacWhisper's one-time fee is the safer, cheaper bet.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Code source gratuit et ouvert sur GitHub ; l'application packagée disponible sur wordmate.sh a un prix non divulgué dans le dépôt
Pros
100% on-device, aucun audio envoyé dans le cloud
Nettoyage automatique des hésitations et faux départs par un second modèle local
Latence sub-300ms, traitement quasi temps réel
Code source ouvert et vérifiable
Cons
Réservé aux Mac Apple Silicon (pas de Windows, Linux ni iOS)
Prix de l'application commerciale non communiqué dans le dépôt
Taux de nettoyage (79-91% mots propres) encore perfectible et auto-déclaré, non vérifié indépendamment