MacWhisper, native macOS app for Whisper

[jordibruin]

First of all, a massive thanks to @ggerganov for making all this! Most of the low level stuff is voodoo to me, but I was able to get a native macOS app up and running thanks to all your hard work!

MacWhisper lets you run Whisper locally on your Mac without having to install anything else.

Features

• Easily record and transcribe audio files
• Just drag and drop audio files to get a transcription
• Get accurate text transcriptions in seconds (up to 15x realtime)
• Search the entire transcript and highlight words
• Supports multiple languages (fastest model is English only)
• Copy the entire transcript or individual sections
• Reader Mode

MacWhisper is very basic right now, so please let me know if you run into anything. You can download it for free here:
http://goodsnooze.gumroad.com/l/macwhisper

[jordibruin]

Added a bunch more features such as editing and deleting segments, as well as language selection. Love this framework, thanks again!

[minfuon96]

I've been using the free version of the MacWhisper app for about a month, and I must say, I'm a huge fan of the "Speaker Paragraph" export function. It's been incredibly helpful in helping me grasp the overall message conveyed by the speakers. I've been contemplating getting the full license, especially because I have a hefty library of over 500 GB of videos that need transcribing.

However, there's one aspect that's been bothering me. When it comes to bulk transcribing videos, I can't seem to find the "Speaker Paragraph" export option anywhere. This restriction has left me wondering why such a limitation exists. Wouldn't it make sense to unlock this export option for paying users who've already invested in the app?

Now, I find myself in a dilemma. I can either manually transcribe videos one by one, ensuring I get the "Speaker Paragraph" export, or I can opt for the PRO license to bulk transcribe my videos, but they'll come out as fragmented sentences.

For me, the absence of the "Speaker Paragraph" export option for bulk-transcribed videos is a significant drawback, preventing me from making the purchase I've been considering for over two weeks now. I kindly request adding the "Speaker Paragraph" export option for both single video transcriptions and bulk transcriptions. Your consideration would be greatly appreciated. Thank you.

[Duseylicious]

I've already purchased and love it, but I'd 2nd this request. 👍

[mrienstra]

@jordibruin, ping, just in case you missed this thread.

[kjkunkle]

@jordibruin Can you add a functionality to hide the MacWhisper dock icon ?

Would be nice since I always have it open for the menubar shortcut, but don't like the app cluttering up my CMD+Tab area !

[imthemastah]

I admit I'm still confused by this. I have the pro version and I all see that the useful option is unavailable?

[pbassham]

Love it! Working well.

Here are a few thoughts from some initial use:

1. Would be good to have a way to stop a transcription in progress. Like if you chose a large file and it is going to take too long. I just dropped a 2nd file on top during it working and it crashed it, so I guess that works, lol.
2. Would be awesome to see a list of transcription history in the sidebar, for instance.
3. Some view preferences would be nice.
   • show/hide timestamps (maybe just show them on the right to save space, or just on hover)
   • reader view by default? (the default view is quite inefficient on space)
   • An editable reader mode seems more useful than the very short segments in the default view right now. That would allow you to group it into paragraphs, separate by person, etc, more easily.
4. Showing the local audio file that is transcribed with a player would be awesome. And clicking on a sentence could jump to that timestamp in the audio file.

Definitely a good start!

Any way to see when updates are released?

[matsumurae]

This is a really good app! I remember last time I tried something similar and the AI with Japanese was a bit shitty… I need to review the extracted text but it looks really good (at least what I've saw on the first mins).

I have two ideas that could be good to have:

1. Would be possible to differentiate between speakers? Well, if not, it's still possible to add it by hand but that's hard work too.
2. I would like to have the option to re-open files. This could be with a proprietary file format or using the exported ones (like .csv or .srt).

About the first point
This could be really useful to transcribe podcasts with multiple speakers.

For the 2nd point
I'm not sure if it's possible to re-open a file without a proprietary format so it stores both audio + text. I'm thinking on something like oTranscribe does with it's own files so you can review as many times as you want.

Also, I would like to ask why MacWhisper isn't on Github or Gitlab so all of us can collaborate too.

[taivlam]

For your first point, Discussion #450 involves what's called "diarization" in whisper.cpp.

Issue #489 (for 3/more speakers) is closed, though really issue #64 is the issue to keep tabs on with respect to diarization progress.

[kai-shimada]

For your second point, we have added support for saving and reopening from SRT files in WhisperScript Pro. We were actually motivated to implement this feature by your suggestion :)

You can see other features of WhisperScript Pro here: https://getwavery.gumroad.com/l/whisperscript

[matsumurae]

For your second point, we have added support for saving and reopening from SRT files in WhisperScript. We were actually motivated to implement this feature by your suggestion :)

Wow, happy to see it was a useful feature.

You can try WhisperScript right here: https://getwavery.gumroad.com/l/whisperscript

I couldn't resist to test it. I've sent to the mail all the feedback that I've found :)

[stephando]

Love it! And works really well.
However the export transcript button does not work on my machine (MacOS Monterey on MacBook Pro with M1 Max) - any thoughts?

[LtPierrot]

Same issue... cannot export

[ReeceKidd]

Same

[SWM89]

Nice job ! Looking further for live transcript…

[dmuiX]

would be nice if i could choose the language

[Sogl]

Nice app. Please add support for *.ogg files (WhatsApp, Telegram voice).

[Jeff9387346]

Fantastic work.

Question: I initially downloaded the non-pro version and chose only the english dataset. After testing I want to try the multi-language dataset, but I see no way in the app or website to go back and get that file. Pointer would be appreciated.

[jmantn]

I purchased MacWhisper Pro and love the application.

I primarily use this for podcasts and hour long recordings. I'd like to request a way to export a transcript and its accompanying audio or video and have it be interactive like click on three paragraphs down and there's where the media will jump to. Or A program that could do that as I'd happily pay for something that like that.

Would also love to see macwhisper come to iOS as well. Gladly pay for this again just to have it on mobile as well.

[Jeff9387346]

Also, once I identify a speaker, it would be fantastic for the AI to then label all instances of that voice appropriately!!

[jmantn]

Also, once I identify a speaker, it would be fantastic for the AI to then label all instances of that voice appropriately!!

That's honestly what I'm hoping for because the alternative would be tedious for longer transcripts.

[canoben]

Batch mode does not work in the Pro version (only reason why I bought it).

[jmantn]

Batch mode does not work in the Pro version (only reason why I bought it).

I'm using hour long .mp3's and I've done up to 5 at a time with the Large model selected and they all worked for me. One nice thing about batch mode that's missing from doing individuals for me is when I do batch mode I can select ahead of time multiple formats for output whereas when I do it individually I have to manually export with each format I want one at a time.

But at least for me and my use case batch mode is working. Might be helpful to the dev to know which macOS version you're running and what file types your trying to convert?

[canoben]

I'm using the latest versions of everything (MacWhisper 2.14 and macOS 13.3.1) and tried with both mp3 and wav files. I've tried on both Intel and Apple Silicon machines, but the batch always stops after the first file (and nothing is written to the disk). Changing language models or output options does not change this. Starting a batch with 5 files or less doesn't help, either.
But I've found a solution for me: I compiled Whisper.cpp myself and use it with the command line. Works perfectly, although strangely much slower than MacWhisper. Maybe I missed some optimisation flags for Apple Silicon. MacWhisper runs much faster on AS compared to the Intel versions.

[dchad3]

I used Mac Whisper Pro Medium to transcribe an interview - stereo file. I wish I could get MW to identify each track with a name. It's too much work to do that manually. I see I can add people but I didn't figure out how that worked. Would be nice if there were YouTube videos showing what it can do and how to do it. Maybe there are but I haven't found them yet.

[dchad3]

Turns out that either I did it wrong or that it doesn't do stereo files. It only transcribed the first track. I made it mono and transcribed it that way and now have to separate the host and guest comments. I don't see how to have host and guest each with separate paragraphs and identified as in Host: blah bla, Guest: blah blah - except manually. Looking forward to Mac Whisper being able to do that.

[dchad3]

I got Mac Whisper for my Macbook Air M1 Ventura and got the pro. It then gave me size options so I chose medium. Can I have Pro large? There was no option to do that with the Pro. I do podcasts with a different person each time and am in no hurry so the best I can get is what I want.

[j-f1]

You should be able to go to the dropdown in the top right corner labelled “Medium” and then choose “Manage Models…” to download additional models.

[dchad3]

Thanks. I did look into that and see for English only, Pro medium seems to be the best. Large has multiple languages.

[wanderingstan]

Just found a little bug: Looks like it only listens to the left channel of a stereo file.

I kept getting only "[BLANK_AUDIO]" for an mp3 file that clearly had voices in it. Mystery was solved when I opened it in audacity and saw that the speaking was all in the right channel.

[iandundas]

Thanks for the feedback - this is fixed in 5.7, which will be available soon, likely within the next week.

[alexgoodell]

There is still something buggy going on with left/right management - I continued to get [BLANK AUDIO] when submitting audio on the L or R (technically it's an audio interface with 4 ins, so this was channels 1 and 2). When I switched to the mic on my webcam it worked perfectly.

EDIT: Dev already fixed for multichannel audio inputs... I just hadn't dug through settings enough. For the free version nonetheless. Seriously incredible. Purchased pro version immediately after seeing this!

[klh]

you should add a settings section to tweak num of cpus / threads weights etc.
should be easy to add.

[auroraflux]

This app is absolutely phenomenal, I've bought it 3 times now for myself and two friends.

Is there any chance there's an equivalent for Android? I'm looking for something that allows at the very least a long, uninterrupted mic recording where I can just put the phone down and let it pick everything up and transcribe it later.

I realize I could just do an audio recording and move the file and transcribe it via the MacWhisperer app later, but if I could do it all on-device that would be incredible.

Thanks!

[ef-is]

[auroraflux]

Or, he could do as the MIT license for OpenAI's Whisper specifically allows him to do:

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so

This is how the MIT open source license works. He is not only well within his rights to charge for it, he should be free to. This is what the spirit of open source licenses are, and this is what software freedom looks like.

If you don't understand how licenses like these work, maybe refrain from from making soapbox-style commentary about it.

[readtedium]

Putting a nice interface on a script and charging for more advanced features is additional work. You are paying for the ease of use—and the dev is nice enough to make most of it available to people for free. If you don’t like this approach, use one of the other tools out there that won’t charge you at all.

[sandipb]

I rarely pay for opensource wrappers unless they provide a significant value add. This gives me a native gui with additional support for managing models based on my requirements, keeps the model list updated, handles different kinds of inputs, allows me to edit transcriptions, etc all value adds to the option of doing all this manually even if I have the tech knowhow to do it myself.
I am completely ok with paying for this.

Now the way you are looking at pricing is backward - he could have charged for every aspect of the app, but customer acquisition strategy is to provide the lower quality fast models for free. It is pretty common paradigm in AI apps right now dont you think? perplexity is doing just the same.

[ef-is]

Or, he could do as the MIT license for OpenAI's Whisper specifically allows him to do:

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so

This is how the MIT open source license works. He is not only well within his rights to charge for it, he should be free to. This is what the spirit of open source licenses are, and this is what software freedom looks like.

If you don't understand how licenses like these work, maybe refrain from from making soapbox-style commentary about it.

you're absolutely right. my bad.

[auroraflux]

Hey, it's always a fair point to bring up, and I can understand feeling a bit indignant about something like this. But it's okay! We learn about these licenses and then we re-evaluate our stance when we get new information.

Don't beat yourself up.

[sandipb]

I am using a paid Pro version of the app. Is this thread the only community for the app? I see a feedback email address for the app but no community links. I really think the app would benefit from either a forum/discord. Even a github discussion is fine.

[sichtbeton]

I am wondering the same thing. Did you find any forum / hub yet?

[johnmarshall4]

I vote for discord, that is where most of the other indie AI apps I use seem to be these days.

[sandipb]

Now that I have marked speakers for the transcription, I see now output format which preserves this information. Exporting segments to pdf or html seems to just export the original timestamp based segments. Is there a way I can add this speaker information to audio segments anywhere?

[bwoodcock]

Hi. I've got a 34-second-long wav file, single speaker, in French, but only the first second gets transcribed... Anyone have any suggestions of what I might do to get this working? There's no apparently difference in the audio or file at the one-second mark, the transcription literally stops in the middle of a sentence, while the audio continues uninterrupted.

Thanks for any ideas!

[paulf12345]

Any suggestions for optimizing speed on an M2 Max MBP 96 GB? It'd be sweet if theres any benchmarks or you have any advice on models (distilled, turbo, normal?), audio pipeline and encoding/decoding format to use (assuming whisperkit is recommended), and effects of pipeline and encoding/decoding compute units on flash attention, greedy vs beam search, etc...

I saw some notes on CoreML + GPU processing in discussions back in march but have not been following repo closely enough to know whether this has been implemented (at which point I assume whisperkit is no longer best option? although tbh idk what the difference between whisperkit and .cpp models are other than swift support).

[johnmarshall4]

For what it's worth, I have a similar machine and I have found the Small English C++ model to be the best speed vs. accuracy. I use it for nearly everything I do. I use flash attention. If you want / need a large model, the 1.5G distilled whisper kit models are faster than Large English C++, but I don't find them as accurate. The non-distilled large whisper kit models have better accuracy but are much slower than C++. At least that is what I've found on an m3 Max. YMMV.

[realAbitbol]

Will ollama be supported as an alternative to openai and anthropic at one point ?

[tembalanco]

Hi, I've been using the new version of Macwhisper (Paid account) for a few weeks. Yesterday I went to launch it on my Macbook M2 and got a crash report. Here is a GTP summary:

Any one run into this? Would appreciate any guidance. Thanks.

---

From the crash report you provided, the key issue seems to be a segmentation fault (EXC_BAD_ACCESS) with SIGSEGV, which occurs when the program attempts to access memory that it shouldn't.

In your case, it looks like the crash is happening in Thread 4, where a function in the MacWhisper app is attempting to read from an invalid address (far: 0x0000000000000004), which isn't mapped to any memory region.

Key Indicators:
KERN_INVALID_ADDRESS at 0x0000000000000004: This indicates an attempt to access memory at an invalid location.
The error is happening during interaction with Swift Concurrency. The function completeTaskWithClosure in libswift_Concurrency.dylib and dynamic casting in libswiftCore.dylib are likely involved, suggesting that the issue may stem from improperly handling async tasks or concurrency.

Potential Causes:
****Swift Concurrency Issue: There may be an issue with the way concurrency is handled in the app, especially around casting or dynamic type checking.

• Memory Management: This could be caused by dereferencing a null pointer or an uninitialized variable.

• External Dependencies: The crash involves YbridOpus and YbridOgg frameworks. There may be an issue with how these frameworks handle concurrency or memory on macOS 15.

• M2-Specific Behavior: Given that the app was working before, it’s possible that the crash is triggered by differences in macOS 15 on Apple Silicon M2, especially related to memory addressing.

Recommended Actions:

• Check Async Code: Review any concurrency or asynchronous task handling in the code, particularly around closures. Ensure proper synchronization when working with shared resources.

• Validate Casting: Review any dynamic casting operations, especially where you are casting between types in Swift. Make sure all optional values are properly unwrapped.

• Check Dependencies: Make sure the YbridOpus and YbridOgg frameworks are fully compatible with macOS 15. You might need to update them to the latest version.

• Test on Another Version of macOS: If possible, test the app on an older version of macOS or on an Intel Mac to see if the issue is specific to macOS 15 or the M2 architecture.

[endlesssmoko]

mine is also crashing. Paid version. Mac M1 with 16gb of ram. I am reinstalling it and will see it that works. I wanted to batch about 150 files and I think that broke it...

[usizu]

Has anybody figured out any way to automate transcriptions with Mac Whisper? It appears to not have AppleScript or Shortcuts support, but I need to do weekly transcriptions and it would be great to be able to control Mac Whisper programatically so it can be part of an automatic workflow.

[maxbourke]

I came here to feature request CLI execution so I can call macwhisper from python or terminal. But Applescript would also work. I believe I could install whisper and torch but it seems a waste to double up model storage.

[stevef243]

I absolutely love the dictation feature it is much better than the Apple dictation but at present the custom shortcut button doesn't work and I also noticed sometimes it makes a dictation noise but the microphone won't pop up and it does not work for instance right now so I have to use Apple dictation which is way less accurate. Any suggestions?

[jd2476]

I absolutely love the app, and that it does the transcription fully offline. @jordibruin, thank you and to everyone who's contributed and help make this app happen.

I get a lot of voice notes which are many times long, and end up feeding them through MacWhisper to transcribe. It would be nice to be able to control the translation feature... I've noticed however that with transcriptions that are bi-lingual (e.g. English-Arabic), some voice notes get transcribed without translation (i.e. English gets transcribed in English, and Arabic in Arabic), however in most cases MacWhisper will translate to one of the languages and only transcribe in that language.

Is there any way to disable translation for a particular transcription, but keep the multi-language detections and transcriptions?

Thanks!

[KitaitiMakoto]

@jordibruin seems not here but at X: https://x.com/jordibruin