MacParakeet - Free + Open-source WisprFlow alternative that runs on Mac Silicon
Posted by PrimaryAbility9@reddit | LocalLLaMA | View on Reddit | 12 comments
I'm on a journey to replacing my monthly SaaS subscriptions. First stop is WisprFlow.
So I built MacParakeet (MacOS only) as a replacement. It's free and open-source under GPL!
I mainly focused on the things that I need, which boiled down to:
- WisprFlow-like UIUX for dictation (smooth + polished)
- YouTube transcription & export to multiple formats
There are some additional features I added, like chat with youtube transcript (integration is available with local ollama or cloud vendors like openai or claude). It runs on NVIDIA's Parakeet model (0.6B-v3) via FluidAudio, which has the best performance for realtime transcription for English. 60 min of audio transcribes in <30 seconds (after the local model has been loaded the first time ofc). WER is also very low.
There are many other similar apps out there with much wider array of features, but I made this for myself and will continue iterating in the spirit of "there are many dictation/transcription apps, but this one is mine." (homage to badlogicgame's pi agent)
How it works
- Press a hotkey in any app, speak, then text gets pasted
- File transcription: drag-drop audio/video files
- Transcribe YouTube URLs via yt-dlp
- Speaker diarization - identifies who said what, with renameable labels
- AI summaries and chat - bring your own API key (OpenAI, Anthropic, Ollama, OpenRouter)
- Clean text pipeline - filler word removal, custom words, text snippets
- Export formats - TXT, Markdown, SRT, VTT, DOCX, PDF, JSON
Limitations:
- Apple silicon only (M1/M2/M3/M4 etc)
- Best with English - supports 25 European languages but accuracy varies; No broad multi-lingual support, so it won't transcribe korean, japanese, chinese, etc.
This app has been in production for about 3 weeks now with 300 downloads thus far. Most of the discovery coming in from organic google search. I've been continually fixing and refining. In any case, I have cancelled subscription to wisprflow (which is a great app and has served me well for many months); local asr models (like Parakeet) and runtime (like FluidAudio) have gotten way too good to ignore.
Hope you like it - let me know!
Website - https://www.macparakeet.com/
Github - https://github.com/moona3k/macparakeet
frogstomper7@reddit
Awesome app! Any plans to add watch folder functionality?
bloknayrb@reddit
Here's my windows version, also FOSS: Talkie
NoFaithlessness951@reddit
you do know that handy also works on macos right?!
PrimaryAbility9@reddit (OP)
yepp, handy is great and all folks should check it out - https://github.com/cjpais/handy
Vicar_of_Wibbly@reddit
Handy runs on Mac, too.
With your stack, what's the limit on how long you can speak before transcription cuts off?
koloved@reddit
https://handy.computer seems better than that cause cross-platform
SatoshiNotMe@reddit
Hex is my current STT app for near-instant transcription with parakeet V3 on my M1 MacBook.
https://github.com/kitlangton/Hex
Uses the same tech stack as this (FluidAudio etc). I’ll see how this compares.
3dom@reddit
GPL is quite limiting license when it comes to business use. Apache 2.0 maybe?
BP041@reddit
been waiting for something like this -- WisprFlow is solid but the subscription for what is essentially a STT wrapper always felt hard to justify.
how does latency compare on M2/M3? whisper.cpp with medium.en gets to around 2-3s on my machine which is acceptable but not seamless for dictation mid-thought.
the YouTube transcription is a nice addition too. that's a separate use case most dictation tools ignore but it's actually where i spend more time -- research notes, reference summaries. good call including it.
PrimaryAbility9@reddit (OP)
The memory footprint for parakeet is much lighter than whisper & I found parakeet's speed to be extremely fast (and overall satisfied with the accuracy). WhisperKit made integration easy when whisper was the hot new model (and deservingly so; whisper was remarkable). Now FluidAudio is doing the same for parakeet. The model is well optimized for apple silicon so the performance/experience is great.
> the YouTube transcription is a nice addition too. that's a separate use case most dictation tools ignore but it's actually where i spend more time -- research notes, reference summaries. good call including it.
Thanks. Youtube website does have its own transcript, but the quality is often poor and I wanted to extract high quality transcription for valuable raw sources.
Btw, there is also a CLI support so you can have your coding agent use it for transcription tasks as well. I recently had claude download \~50 tiktok videos and ran transcription via CLI (sequential processing).
BP041@reddit
the CLI + agent integration is the part I wasn't expecting but makes total sense in hindsight. once you have a local transcription binary that just works, any research agent can pipe through it without API costs or privacy concerns -- and sequential CLI processing is fine for batch jobs where you're not waiting in real-time.
the 50 TikTok batch is actually the more interesting use case to me than live dictation. I've been doing something similar with long podcasts -- grab audio, transcribe locally, then have the agent extract key claims rather than sitting through the whole thing. the bottleneck is usually the "get the audio" step, not the transcription itself.
the memory footprint point changes my calculus though. Whisper medium.en at 2-3s I can live with, but it's competing for ANE with everything else running. if Parakeet is lighter at comparable accuracy that's the real reason to switch, not just the latency number.
germanheller@reddit
nice, parakeet is surprisingly good for its size. ive been using whisper ONNX models (tiny/base/small via @huggingface/transformers) for dictation in an electron app and the latency after initial model load is under 400ms on most machines.
curious about the FluidAudio integration -- does it handle streaming input or does it batch process after you stop talking? thats the main UX difference that makes or breaks dictation tools imo. wispr feels instant because of the streaming, most open source alternatives feel sluggish because they wait for silence