Feature Request: also provide models in torchscript.

[SR-Lut3t1um]

Request in title.

Would be really helpful to integrate WhisperSpeech easily into e.g. android.

[zoq]

Technically this should be possible, I marked this as a good first issue, in case someone likes to take a closer look into this.

[jpc]

It may be worth looking into the recent ExecuTorch as a possible alternative.

WhisperSpeech already works with torch.compile but I think we had some issues with torch.export around our implementation of RoPE.

[BBC-Esq]

This also worked:

ffmpeg -hide_banner -loglevel panic -i "test 1.mp3" -threads 1 -acodec pcm_s16le -ac 1 -af aresample=resampler=soxr -ar 16000 "test.wav" -y

[SR-Lut3t1um]

It seems to be that Executorch isn't compatible with Androids NNAPI. However I'm unsure if that leads to worse performance.

[gongouveia]

Hello, @SR-Lut3t1um is exporting as .onnx a good alternative?

[SR-Lut3t1um]

Well, according to the pytorch docs torchscript is required, so I'm afraid. onnx isn't going to be an alternative for that use case.

[gongouveia]

I am pretty sure it is possible to use ONNX Runtime and the NNAPI execution provider.
perhaps converting to ONNX may be less of a hassle

[SR-Lut3t1um]

Yeah onnx should be fine too