FnKey
macOS dictation with Deepgram stream
103 followers
macOS dictation with Deepgram stream
103 followers
A tiny Rust menu bar app for macOS. Hold Fn to activate the mic, speak, release to paste. Audio streams to Deepgram Nova-3 in real time — no batch delay. Falls back to Groq Whisper. Open source, free.
FnKey
Been using Deepgram's streaming API for call-center transcriptions - the latency is wild. Curious how FnKey handles overlapping speech when you're dictating code while thinking aloud? My current workaround is pausing like a Victorian telegram, which defeats the purpose of stream processing.
FnKey
@lliora With Deepgram, I don't see latency. It's actually lightning fast.
Maybe it's a new version of their model. Compared with what you've been working with.
There is no overlapping handling - I never experienced it myself, honestly.
Interesting take on trying to optimize on latency. I would be more interested in how dictation tools handle punctuation, formatting text, removing filler words, etc. Congrats on launch!
FnKey
@tteer Somehow, the Deepgram is pretty good. Punctuation is built in their Speech-to-Text.
LiveDemo
Hi @evoleinik
This sound like a very useful app,
Do you guys plan to support importing voice from audio stream, like from Teams or Slack?
Because let's say my team speaks French or Portuguese,
I just click the button and it would start streaming my team audio to the transcription service
And it would translate almost instantly
I know that teams supports something similar, but it is not flexible in terms of models,
And if you guys stream it directly from the computer audio source, it might be better
FnKey
@gapostolov I think Deepgram only supports English. At least when I speak in different language. It doesn't recognize it.
Looks cool! Does it handle technical jargon well out of the box — model names, CLI flags, that kind of thing? That's usually where dictation tools fall apart i felt
FnKey
@abhinavramesh (1) I feel like voice typing is still good for talking to your AI. For talking to people it's not there yet. For talking to people we may want to add light post-processing with a fast model - I experimented briefly with it, but it was messing with the sentence structure too much.
(2) Deepgram worked better than Whisper in practice for me. Still, when I see misspelled term, I just add it to the list of custom keywords:
# Custom keywords (one per line)
Anthropic
Claude
AirShelf
claude-grep
grep
Haiku
OpenClaw
Vercel
Claude Cowork
Mullvad
cron
cronjob
crontab
Sid
pandoc
BrightData
BIME
Tuco
12Go
neon
fnkey
FnKey
@sarrah What key would you like? I can set it up.
Built in Rust, streaming to Deepgram over WebSocket, and open source. That's a solid stack. How accurate is the transcription for technical jargon or code related terms like function names and variable names? Would love to use this for dictating code comments. Nice work on shipping this!
FnKey
@carmonamar51249 I think we need a way for it to somehow see your current context. Otherwise, it will be guessing.