Ben Nortier (@bjnortier) profile picture

Creator of Hello Transcribe. Software Engineer living in Cape Town.

Earth
Joined March 2009
Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)Photo by Ben Nortier (@bjnortier)
Hello Transcribe 5 is now in Beta 🙌 • NVIDIA Nemotron 3 Diarization • Core AI NVIDIA Parakeet v3 • iPhone Duo Support • Everything local • … and much more More info and sign up for the Beta here: bjnortier.com/articles/hello… Demo recorded on iPhone 15 Pro
2
216
Muse Spark 1.3 isn’t down
86
Something I’ve been working on… Meta’s SAM 3D Body working on-device with Apple Core AI 🙌 Inference performance: 0.5FPS on an iPhone 15 Pro 5FPS on Mac M5 Pro
1
1
231
Inference videos
1
42
I tried Gemma4 E4B to analyse security camera footage for suspicious activity… “Based on a review of the entire video footage, there is **no suspicious activity** observed.” :/ bjnortier.com/articles/gemma…
1
1
1
518
I built Hello Transcribe because I hate listening to Voice Notes. WhatsApp now has transcriptions but they aren't great and who knows where your data is sent? Here's how to use Hello Transcribe to transcribe Voice Notes accurately and on-device: bjnortier.com/blog/transcrib…
1
104
Hello Transcribe version 4 is out 🥳 - Large v3 Turbo support and more free models - Audio Player and Karaoke text view - Audio Persistence in PRO - iOS26 & UI updates - Recover from failed dictation Read more here: bjnortier.com/blog/hello-tra… App Store: apps.apple.com/za/app/hello-…
1
3
259
I love programming in #Swift
3
449
Here's the blog post about Hello Transcribe 3.2: - 400% speedup. - LargeV3 support. - Fast aborts. bjnortier.com/2023/11/17/Hel… #AI #OpenAI #Whisper
2
8
594
Time reductions
3
1
312
I explain I recommend just using beam=5 by default and choosing a bigger model for more accuracy (I will add a new guide to the app), and using LargeV3 over LargeV2:
1
1
304
Hello Transcribe 3.2 is out: apple.co/49eTzRn With beam search size of 5, and 30s of audio, transcription times are down as follows: Mac M1 Max: tiny: -79% base: -65% small: -76% medium: -84% large-v2: -85% iPhone 14 Pro: tiny: -64% base: -61% small: -70% medium: -85%
3
2
15
3,962
I recommend using a beam of 5, reducing it will give a speed improvement but you will lose accuracy. Blog post coming later today. #AI #OpenAI #Whisper
2
270
Updated graph for Whisper.cpp 1.5.0 speed improvements which will be in the next version of Hello Transcribe. Beta out today and hopefully submitting to App Store later today. Basically 400% on Mac if you use beam search of 5. #AI #OpenAI #Whisper
2
31
4,616
The improvement in performance with @ggerganov’s Whisper.cpp+Metal is massive for the bigger models, but equally important is the improved usability for the smaller models and CoreML. It’s very difficult to recommend whether to use CoreML — … 1/2
1
5
459
the caching step can lead to unexpected and uncatchable crashes, can take forever (17mins on macOS for small), and the OS can remove the cache when it decides to. Now things will be much simpler for users 🙏 2/2 #AI #OpenAI #Whisper
1
239
Massive speed increase coming to Hello Transcribe 3.2 for Medium and Large thanks to @ggerganov & Whisper.cpp support for Metal. Beta out today (DM me for access). Also: - Support for new Large V3. - Fast abort. - Support beam search size == 0 (greedy). #AI #OpenAI #Whisper
1
5
1
60
13,494
You can now transcribe videos from the Photos app (.mp4 and .mov) in Hello Transcribe 3.1. Also fixed a significant memory leak. Get the update here: apple.co/49eTzRn #AI #OpenAI #Whisper
6
1,443