I'll have you know that I'm Matt's top contributor to Ghost Pepper and I'm nearly fifty
But I did it because I wanted it to work exactly the way I wanted it.
Also, for kicks, I (codex) ported it to Linux. But because my Linux laptop isn't as fast, I've had to use a few tricks to make it fast. https://github.com/obra/pepper-x
On Linux, there's access to the latest Cohere Transcribe model and it works very, very well. Requires a GPU though. Larger local models generally shouldn't require a subordinate model for clean up.
Have you compared WhisperKit to faster-whisper or similar? You might be able to run turbov3 successfully and negate the need for cleanup.
Incidentally, waiting for Apple to blow this all up with native STT any day now. :)
How does it compare to the more well established https://github.com/cjpais/handy? Are there any stand out features (for either option)? What was the reason for writing your own rather than using or improving existing software?
I've been running whisper large-v3 on an m2 max through a self-hosted endpoint and honestly the accuracy is good enough that i stopped bothering with cleanup models. The bigger annoyance for me was latency on longer chunks, like anything over 30 seconds starts feeling sluggish even with metal acceleration. Haven't tried whisperkit specifically but curious how it handles longer audio compared to the full model.
Yeah that makes sense, chunking on silence would sidestep the latency issue pretty cleanly. I've been running it through a basic fastapi wrapper so it just takes whatever audio blob gets thrown at it, no chunking logic on the server side. Might be worth adding a vad pass before sending to whisper though, would cut down on processing dead air too.
Whisper is still old reliable - I find that it's less prone to hallucinations than newer models, easier to run (on AMD GPU, via whisper.cpp), and only ~2x slower than parakeet. I even bothered to "port" Parakeet to Nemo-less pytorch to run it on my GPU, and still went back to Whisper after a couple of days.
Thank you for sharing, I appreciate the emphasis on local speed and privacy. As a current user of Hex (https://github.com/kitlangton/Hex), which has similar goals, what are your thoughts on how they compare?
I see quite a few of these, the killer feature to me will be one that fine tunes the model based on your own voice.
E.G. if your name is `Donold` (pronounced like Donald) there is not a transcription model in existence that will transcribe your name correctly. That means forget inputting your name or email ever, it will never output it correctly.
Combine that with any subtleties of speech you have, or industry jargon you frequently use and you will have a much more useful tool.
We have a ton of options for "predict the most common word that matches this audio data" but I haven't found any "predict MY most common word" setups.
Handy is an awesome project, highly recommended - many of our engineers and PMs use it! CJ, Handy's creator, recently joined us as a Builder in Residence at Mozilla.ai. So for those interested in deploying a more raw/lightweight approach to local speech-to-text (or other multimodal) models, feel free to check out llamafile - which includes whisperfile, a single-file whisper.cpp + cosmopolitan framework-based executable. We're hoping to build some bridges between the two projects as well. https://github.com/mozilla-ai/llamafile
I’d also be interested to know what the impetus was for developing ghost-pepper, which looks relatively recent, given that Handy exists and has been pretty well received.
Extra bonus is that Handy lets add an automatic LLM post-processor. This is very handy for the Parakeet V3 model, which can sometimes have issues where it repeats words or makes recognition errors for example, duplicating the recognition of a single word a dozen dozen dozen dozen dozen dozen dozen dozen times.
Yep. Using Handy with Parakeet v3 + a custom coding-tailored prompt to post-process on my 2019 Intel Mac and it's been working great.
Once in a while it will only output a literal space instead of the actual translation, but if I go into the 'history' page the translation is there for me to copy and paste manually. Maybe some pasting bug.
Handy is awesome! I used it for quite a while before Claude Code added voice support. Solid software, very good linux and mac integration. Shoutout to Parakeet models as well, extremely fast and solid models for their relatively modest memory requirements.
I love and have been using handy for a while too, what we need is this for mobile apps I don't think there's any free apps and native dictation is not always fully local and not as good.
Cool, I've been doing a lot of "coding" (and other typing tasks) recently by tapping a button on my Stream Deck. It starts recording me until I tap it again. At which point, it transcribes the recording and plops it into the paste buffer.
The button next to it pastes when I press it. If I press it again, it hits the enter command.
Not sure why I should use this instead of the baked-in OS dictation features (which I use almost daily--just double-tap the world key, and you're there). What's the advantage?
I haven't used this one but WisprFlow is vastly better than the built-in functionality on MacOS. Apple is way behind even startups, even for fundamental AI functionality like transcribing speech
If you don't feel like downloading a large model, you can also use `yap dictate`. Yap leverages the built-in models exposed though Speech.framework on macOS 26 (Tahoe).
I like this idea and it should work -- whatever microphone you have on should be able to hear the speaker. LMK if not (e.g., are you wearing headphones? if so, the mic can't hear the speaker)
Hi Matt, there's lots of speech-to-text programs out there with varying levels of quality. 100% local is admirable but it's always a tradeoff and users have to decide for themselves what's worth it.
Would you consider making available a video showing someone using the app?
Exactly my question. I double-tap the control button and macOS does native, local TTS dictation pretty well. (Similar to Keyboard > Enable Dictation setting on iOS.)
The macOS built-in TTS (dictation) seems better than all the 3rd party, local apps I tried in the past that people raved about. I have tried several.
Is this better somehow?
If the 3rd party apps did streaming with typing in place and corrections within a reasonable window when they understand things better given more context, that would be cool. Theoretically, a custom model or UX could be "better" than what comes free built into macOS (more accurate or customizable).
But when I contacted the developer of my favorite one they said that would be pretty hard to implement due to having to go back and make corrections in the active field, etc.
I assume streaming STT in these utilities for Mac will get better at some point, but I haven't seen it yet (been waiting). It seems these tools generally are not streaming, e.g. they want you to finish speaking first before showing you anything. Which doesn't work for me when I'm dictating. I want to see what I've been saying lately, to jog my memory about what I've just said and help guide the next thing I'm about to say. I certainly don't want to split my attention by manually toggling the control (whether PTT or not) periodically to indicate "ok, you can render what I just said now".
I guess "hold-to-talk" tools are for delivering discrete, fully formed messages, not for longer, running dictation.
AFAICT, TFA is focused on hold-to-talk as the differentiator, over double-tap to begin speaking and double-tap to end speaking?
Thanks! We currently have 2 multi-lingual options available:
- Whisper small (multilingual) (~466 MB, supports many languages)
- Parakeet v3 (25 languages) (~1.4 GB, supports 25 languages via FluidAudio)
But I did it because I wanted it to work exactly the way I wanted it.
Also, for kicks, I (codex) ported it to Linux. But because my Linux laptop isn't as fast, I've had to use a few tricks to make it fast. https://github.com/obra/pepper-x
I have collected the best open-source voice typing tools categorized by platform in this awesome-style GitHub repo. Hope you all find this useful!
https://github.com/primaprashant/awesome-voice-typing
On Linux, there's access to the latest Cohere Transcribe model and it works very, very well. Requires a GPU though. Larger local models generally shouldn't require a subordinate model for clean up.
Have you compared WhisperKit to faster-whisper or similar? You might be able to run turbov3 successfully and negate the need for cleanup.
Incidentally, waiting for Apple to blow this all up with native STT any day now. :)
But in this case I built hyprwhspr for Linux (Arch at first).
The goal was (is) the absolute best performance, in both accuracy & speed.
Python, via CUDA, on a NVIDIA GPU, is where that exists.
For example:
The #1 model on the ASR (automatic speech recognition) hugging face board is Cohere Transcribe and it is not yet 2 weeks old.
The ecosystem choices allowed me to hook it up in a night.
Other hardware types also work great on Linux due to its adaptability.
In short, the local stt peak is Linux/Wayland.
If this needs nvidia CPU acceleration for good performance it is not useful to me, I have Intel graphics and handy works fine.
That said: If handy works, no need whatsoever to change.
Not sure how you're running it, via whichever "app thing", but...
On resource limited machines: "Continuous recording" mode outputs when silence is detected via a configurable threshold.
This outputs as you speak in more reasonable chunks; in aggregate "the same output" just chunked efficiently.
Maybe you can try hackin' that up?
Have you ever considered using a foot-pedal for PTT?
Apple incidentally already has native STT, but for some reason they just don't use a decent model yet.
Apparently they do have a better model, they just haven't exposed it in their own OS yet!
https://developer.apple.com/documentation/speech/bringing-ad...
Wonder what's the hold up...
For footpedal:
Yes, conceptually it’s just another evdev-trigger source, assuming the pedal exposes usable key/button events.
Otherwise we’d bridge it into the existing external control interface. Either way, hooks are there. :)
Parakeet does both just fine.
I've been using parakeet v3 which is fantastic (and tiny). Confused why we're still seeing whisper out there, there's been a lot of development.
It's also in many flavours, from tiny to turbo, and so can fit many system profiles.
That's what makes it unique and hard to replace.
Also vibe coded a way to use parakeet from the same parakeet piper server on my grapheneos phone https://zach.codes/p/vibe-coding-a-wispr-clone-in-20-minutes
E.G. if your name is `Donold` (pronounced like Donald) there is not a transcription model in existence that will transcribe your name correctly. That means forget inputting your name or email ever, it will never output it correctly.
Combine that with any subtleties of speech you have, or industry jargon you frequently use and you will have a much more useful tool.
We have a ton of options for "predict the most common word that matches this audio data" but I haven't found any "predict MY most common word" setups.
https://developers.openai.com/cookbook/examples/whisper_prom...
https://github.com/cjpais/handy
Extra bonus is that Handy lets add an automatic LLM post-processor. This is very handy for the Parakeet V3 model, which can sometimes have issues where it repeats words or makes recognition errors for example, duplicating the recognition of a single word a dozen dozen dozen dozen dozen dozen dozen dozen times.
"You know what would be useful?" followed by asking your LLM of choice to implement it.
Then again for a lot of scenarios it's your slop or someone else's slop.
I think the only difference is that I keep my own slop tools private.
Once in a while it will only output a literal space instead of the actual translation, but if I go into the 'history' page the translation is there for me to copy and paste manually. Maybe some pasting bug.
The button next to it pastes when I press it. If I press it again, it hits the enter command.
You can get a lot done with two buttons.
What makes the others vastly better?
Whisper models I barely bother checking anymore
https://wisprflow.ai/new-funding
I did that so that I could record my own inputs and finetune parakeet to make it accurate enough to skip post-processing.
[0]: https://github.com/beingpax/VoiceInk
Project repo: https://github.com/finnvoor/yap
You could hook it up to some workflow over the local API depending on how you want to dump the text, but the web UI is good too.
The Show HN by the author was at: https://news.ycombinator.com/item?id=44145564
EDIT: I see there is an open issue for that on github
Would you consider making available a video showing someone using the app?
The macOS built-in TTS (dictation) seems better than all the 3rd party, local apps I tried in the past that people raved about. I have tried several.
Is this better somehow?
If the 3rd party apps did streaming with typing in place and corrections within a reasonable window when they understand things better given more context, that would be cool. Theoretically, a custom model or UX could be "better" than what comes free built into macOS (more accurate or customizable).
But when I contacted the developer of my favorite one they said that would be pretty hard to implement due to having to go back and make corrections in the active field, etc.
I assume streaming STT in these utilities for Mac will get better at some point, but I haven't seen it yet (been waiting). It seems these tools generally are not streaming, e.g. they want you to finish speaking first before showing you anything. Which doesn't work for me when I'm dictating. I want to see what I've been saying lately, to jog my memory about what I've just said and help guide the next thing I'm about to say. I certainly don't want to split my attention by manually toggling the control (whether PTT or not) periodically to indicate "ok, you can render what I just said now".
I guess "hold-to-talk" tools are for delivering discrete, fully formed messages, not for longer, running dictation.
AFAICT, TFA is focused on hold-to-talk as the differentiator, over double-tap to begin speaking and double-tap to end speaking?