Now that's useful for bypassing the regular Android transcription that (tries to) send the audio to some Google servers.
Is it possible to support dropping in custom compatible models? Maybe transformed ones like for faster-whisper? That could also save the download for already downloaded models on the PC.
The google record app (and the old leaked that works on other androids) use local model too.
I mean this is great of course. Just saying not everything needs to be sent to google
But what is the advantage? If you have a German Tiny model with 75MB and I have a multi-lingual base model with 78MB? Is the German tiny better than multi-lingual base?
The advantage is that a model specifically tuned for a language, like the one that I linked, provides substantially better transcription at the same model size, well, or faster transcription at the same quality, which is nicer for mobile devices.
I just managed to convert your model. The tflite has 42 MB. But in a first test it is much worse than the multi-lingual base model I have. Of course it is about twice as fast.
I usually use the small model. It is much slower but usually gives perfect transcription which does not need any manual editing afterwards...
in case you manage to convert it to tflite such that it is working with my app please open a pull request for my Huggingface tflite repo. Then others might be able to use your model as well.
Thanks, I also just found that one while following links :-)
Having this automated "put what you need in at the top, and you get something that works with the app at the bottom" is great to have.
Thanks, I just finished setting this duo up. And because it wasn't obvious to me, if Heliboard keeps defaulting to Google Voice Typing, you have to disable it in Android's list of keyboards so that Whisper is the only active voice option. (I kept looking for voice typing settings in HeliBoard's settings.)
i use HeliBoard. have selected both input methods. you can use Heliboard's microphone button to switch to whisper. and back to the KB with the back button
Chromix_@reddit
Now that's useful for bypassing the regular Android transcription that (tries to) send the audio to some Google servers.
Is it possible to support dropping in custom compatible models? Maybe transformed ones like for faster-whisper? That could also save the download for already downloaded models on the PC.
FPham@reddit
The google record app (and the old leaked that works on other androids) use local model too.
I mean this is great of course. Just saying not everything needs to be sent to google
DocWolle@reddit (OP)
But what is the advantage? If you have a German Tiny model with 75MB and I have a multi-lingual base model with 78MB? Is the German tiny better than multi-lingual base?
Chromix_@reddit
The advantage is that a model specifically tuned for a language, like the one that I linked, provides substantially better transcription at the same model size, well, or faster transcription at the same quality, which is nicer for mobile devices.
DocWolle@reddit (OP)
I just managed to convert your model. The tflite has 42 MB. But in a first test it is much worse than the multi-lingual base model I have. Of course it is about twice as fast.
I usually use the small model. It is much slower but usually gives perfect transcription which does not need any manual editing afterwards...
DocWolle@reddit (OP)
in case you manage to convert it to tflite such that it is working with my app please open a pull request for my Huggingface tflite repo. Then others might be able to use your model as well.
DocWolle@reddit (OP)
you can also use other models if they are .tflite and have the right signatures.
I am using this Colab for conversion: https://huggingface.co/DocWolle/whisper_tflite_models/blob/main/Generate_tflite_for_whisper_base_with_transcribe_and_translate_signatures.ipynb
You need to copy the model to Android/data/org.woheller69.whisper/files
If your phone does not allow that you need to use adb push from PC.
Chromix_@reddit
Thanks, I also just found that one while following links :-)
Having this automated "put what you need in at the top, and you get something that works with the app at the bottom" is great to have.
tessellation@reddit
thank you!
this and HeliBoard are my current drivers in the phone.
because it wasn't obvious to me I just wanted to mention that the combination of this two input methods works like a charm for me.
TemperFugit@reddit
Thanks, I just finished setting this duo up. And because it wasn't obvious to me, if Heliboard keeps defaulting to Google Voice Typing, you have to disable it in Android's list of keyboards so that Whisper is the only active voice option. (I kept looking for voice typing settings in HeliBoard's settings.)
Elegast-Racing@reddit
Neat. Works. Would be nice to implement it into the keyboard somehow rather than a separate app to make it useful. Cheers.
tessellation@reddit
i use HeliBoard. have selected both input methods. you can use Heliboard's microphone button to switch to whisper. and back to the KB with the back button
JorG941@reddit
Futo Voice works with other lenguages and works pretty good, and of course its all run locally
DocWolle@reddit (OP)
Futo voice also uses Whisper, so languages are the same
OriginalSimon@reddit
It would be great to see a recorder with built-in Whisper. Google Recorder doesn't support the language I need for transcription.
MikePounce@reddit
Hi, in dark mode the purple text on black background is almost unreadable to me. Checkboxes should be aligned.
DocWolle@reddit (OP)
it can transcribe any language or translate to English