Show HN: Mumble Dictation – local dictation that learns your vocabulary

6 points by tarafans7 ↗ HN
Hey HN! It’s Wen from Narya.ai. We build Mumble Dictation, a local-first and privacy-focused dictation app on Mac.

Homepage and Download: https://heymumble.com/dictation

YouTube:https://youtu.be/Bma59TrR62c?si=DM-LyjZY5GZ9-f64

Two things kept bugging us about the dictation products on the market:

1. Most of them ship your audio to the cloud, and those who is claiming to be private are often just a thin wrapper of cloud APIs with a privacy policy.

2. Those general-purpose ASR models are terrible at your vocabulary: library names, co-workers' names, project codenames, etc.

Mumble Dictation is built around those problems. Its full pipeline including ASR and transcript cleanup is running on your device with local models. Your audio never leaves your machine. Plus, it collects the hot words in your daily usage and gets them spelled correctly.

In particular, the app learns from what you type and what's on your screen, entirely locally, and builds a personal dictionary. At inference time, we boost those terms during decoding, so "Supabase", "PowerSync", or your teammate's name comes out right the first time, instead of some phonetic guess. The dictionary never leaves your machine for sure.

A few other things it does:

- Multilingual (40 languages)

- Custom transcript transform (concise, bullet points, translate, or any prompt you can define yourself)

- Bring your own API key (OpenAI, Anthropic, Ollama, etc.) if you want a different LLM model for transcript cleanup

- Voice shortcuts: say "buy paper towels" and it opens Amazon with the search filled in, or "ask Claude about dictation apps" and it opens Claude with your question already submitted

Pricing is a one-time purchase, no subscription. It's $50 at launch (half the planned price).

Happy to answer questions here.

2 comments

[ 14.3 ms ] story [ 209 ms ] thread
The vocabulary-learning angle is the right problem to chase -- most ASR failures I see aren't model quality, they're proper nouns. Curious how the personal dictionary handles near-miss homophones once it's learned both. Also curious whether you've had bluetooth-mic-specific complaints -- even a great local pipeline is only as good as what the mic hands it, which is the piece we're building hardware around.