3 ms·
My understanding is that Talon is built off of wav2letter's inference framework for ASR.
by whimsicalism 6y ago
My understanding is that Talon is built off of wav2letter's inference framework for ASR.
- lunixbochs 6y agoI do not use the wav2letter@anywhere inference frontend - I trained the acoustic model using the facebook upstream code, but the decoder is almost entirely new, and on Windows I use Pytorch for inference. Talon ships with a libw2l.so/dylib on Linux/Mac built from my open source repos.
- whimsicalism 6y agoInteresting. What model are you using for your acoustic - the streaming convnet? I didn't know that there was a pytorch implementation of the w2l architectures..
- nshm 6y agoFeels like wav2letter will not be actively developed anymore. Understandable since it is hard to compete with Pytorch with a custom NN toolkit. Any plans to move to Pytorch/Tensorflow?
- lunixbochs 6y agoI don’t think that’s right at all. They moved development to the Flashlight repo. It seems very actively developed to me. Last commit 3 hours ago. wav2vec 2.0 blog post also went up last month (September) and the current state of the art iirc is a Google model based on Facebook’s Wav2vec 2.0 work. For my own use I’ve already built Pytorch and CoreML frontends, with a shared model format (can convert models to/from wav2letter format and my custom format), and I have the ability to create new models in these frameworks from wav2letter architecture files. I still run my training in the wav2letter framework, but for compatible training in Pytorch I would mostly just need criterion implementations. I assume warpCTC is fine for the CTC models. There’s also a third party Pytorch ASG criterion package but I haven’t tried it yet.