Does anybody know of a better #speechToText alternative to this?
This feels like a terrible hack that keeps breaking. I decided to look for alternatives after I saw them using /dev/shm to store ML models.
QuantiusBenignus/BlahST
https://github.com/QuantiusBenignus/BlahST
SpeechNote (aka dsnote) does not qualify since it doesn't integrate with the clipboard.
🚀 #Whisperphp Makes Speech Recognition Accessible in #PHP
🔧 New #PHP binding for #Whispercpp brings powerful #AI speech recognition capabilities:
• Supports #Linux (x86_64/arm64) and #macOS platforms with both high and low-level APIs for maximum flexibility
@itsfoss Well, it's probably better to have #WhisperCpp integrated in #Shotcut than to wait until audio exports just to put it through AI externally again.
whispercppがインストールできるようになった
RubyGems.orgにはまだ乗ってないhttps://blogs.kitaitimakoto.net/~/Apehuci/whispercppがインストールできるようになった/
whisper.cppのRubyバインディングが存在していた
https://blogs.kitaitimakoto.net/~/Apehuci/whisper.cppのRubyバインディングが存在していた/
Russian talk radio:
[00:00:00.000 --> 00:00:06.140] Кто должен задать эти новые, что такое хорошо и что такое плохо? Государство?
[00:00:06.140 --> 00:00:16.060] Я думаю, ну, какая-то государственная комиссия, ну, такая реальная комиссия, реальная, которая готова заглянуть в будущее.
[00:00:16.060 --> 00:00:18.500] Кого мы хотим сейчас воспитать?
[00:00:18.500 --> 00:00:23.140] Кого мы хотим воспитать? Я не очень понимаю.
@bert_hubert @hanno +1 #WhisperCpp is well-documented and straight-forwardly setup-able.
Experimented with #whispercpp today. Results are quite impressive. I let it transcribe a 3 min snippet of an interview in German. What I noticed: -medium works significantly better than the smaller models. But it smoothens the text quite a bit, removing duplications, interjections etc., which might be undesirable for academic purposes. The OpenVINO version runs significantly faster even on an Intel GPU, but not by a magnitude. Having everything local is a huge plus for sensitive data.
Seeing an epidemic of people using automatic captioning tools and not actually reviewing the output. Numerous obvious, easily fixed errors.
This does absolutely no favors to people actually _depending_ on those captions to be accurate.
This actually happened a couple weeks ago but I didn't see it until now. They've added experimental "speaker diarization" to #whispercpp. This allows it to mark separations when the person speaking changes. It appears that this uses both the sound of the voice and semantic clues simultaneously.
#machinelearning https://github.com/ggerganov/whisper.cpp#speaker-segmentation-via-tinydiarize-experimental
#whispercpp continues to evolve. :owi: Among many other changes the ones I find most interesting are:
- Multiple transcriptions can run in parallel using a single copy of the model in memory (each using a private state structure but sharing access to the model).
- On many android devices the native 16-bit float (fp16) instructions can now be utilized.
- CoreML can now be used. This is neat but is inherently Apple-only.
https://github.com/ggerganov/whisper.cpp/discussions/766
what happens when you get whisper.cpp to listen to #chiptunes?
both speak at 16KHz so they should understand each other, right?
https://github.com/ggerganov/whisper.cpp
Track: Alpha by @lukhash
Forget about #ChatGPT, I give you #whispercpp #OpenAI https://www.newyorker.com/tech/annals-of-technology/whispers-of-ais-modular-future
Whisper.cpp 1.2 is out with very significant reductions in memory usage with the same models. Roughly in half.
#whispercpp #whisper #speechrecognition #speechtotext
https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.0
Wonder how much € would be saved if people only knew about free and/or open source solutions. #whisperai #whispercpp #subtitleedit
@brainwane @simon Have you looked at #whispercpp yet? I use it daily to catalog my #podcasts and #index the contents. It's WAY faster than normal #python-based whisper...even with #PyTorch. Then again, it's got a limitation on the #algorithms it uses.. it's #inference & #greedy-only right now.
