- need to request the model on hugging face (can't just download)
- not allowed to distribute model without commercial license (so users have to request model themselves? or at least projects that fork or use my open source transcription component?)
BTW I always test with the phrase "Leftium, the element of creativity." Most models fail on "Leftium," but Whistle works if added as a keyword.
---
Here's a test result for you:
Load 0.8 s Clip 3.5 s of speech Transcribe 70 ms (median of 3, beam search width 4, as in the app) Speed 20 ms per second of audio, 50× faster than real time Relaxed SIMD yes
Also if you run the test at the bottom of the homepage, do post it! I've only been able to test on a handful of devices.
e.g for my workstation:
Load
0.9 s
Clip
3.5 s of speech
Transcribe
116 ms (median of 3, beam search width 4, as in the app)
Speed
33 ms per second of audio, 30× faster than real time
Relaxed SIMD
yes
Very impressive~ support for real-time live transcription is great!
I think there is a bug: the settled text seems to fail to get the bold styling.
---
I wanted to try adding to https://rift-transcription.vercel.app, but the project/license terms make it a little cumbersome:
- need to request the model on hugging face (can't just download)
- not allowed to distribute model without commercial license (so users have to request model themselves? or at least projects that fork or use my open source transcription component?)
- no API/documentation
---
Thanks for pointing me to Whistle, which is nearly as small and has simpler project/license terms: https://cactuscompute.com/blog/whistle
BTW I always test with the phrase "Leftium, the element of creativity." Most models fail on "Leftium," but Whistle works if added as a keyword.
---
Here's a test result for you:
Load 0.8 s Clip 3.5 s of speech Transcribe 70 ms (median of 3, beam search width 4, as in the app) Speed 20 ms per second of audio, 50× faster than real time Relaxed SIMD yes
w/ M4 Mac Mini, Chrome
Appreciate the notes and thoughts. Crazy to see how performant the m4 is.
> I think there is a bug: the settled text seems to fail to get the bold styling.
I can't repeat this, the 'bold' is just slightly darker than the actively-worked on words.
I've been noodling on opening the model license further and also the model engine code. Haven't quite figured out what I wanted tbh.
There's a correction mode with starbright - if you correct a transcript, the engine is able to feed those corrections the next go around.
Also if you run the test at the bottom of the homepage, do post it! I've only been able to test on a handful of devices.
e.g for my workstation:
Load 0.9 s Clip 3.5 s of speech Transcribe 116 ms (median of 3, beam search width 4, as in the app) Speed 33 ms per second of audio, 30× faster than real time Relaxed SIMD yes
w/ AMD Ryzen™ AI Max+ PRO 395, firefox
How do you do it? It seems too good to be true based on the comparisons.
Also it's not clear what this means? "25× faster than real-time audio"
It just takes 1 second to transcribe 25 seconds of audio?
yep, averaged out over the length of the audio - as its not immediately always '25x'.
I think the claim is a bit obtuse, might be worth improving the sentence to be more specific
It definitely gave me pause and suspicion
[flagged]