Parcollet Titouan comments

Results 231 comments of


                                            Parcollet Titouan

Multi-GPU issue when pre-training Wav2Vec2

More importantly, are you using layerdrop? Or any form of alteration to the yaml, like LoRA? Are you changing anything?

Multi-GPU issue when pre-training Wav2Vec2

@rajeevbaalwan can you try setting encoder_layerdrop to zero?

add GigaSpeech dataset in SpeechBrain

BON, what must be done here?

add GigaSpeech dataset in SpeechBrain

Taking over this. Training models. I will merge before we have the model, @asumagic will need to train as well as well following the best yaml I provide to replicate.

Encountered a TypeError: expected string or bytes-like object when training the model using dual-GPU model sharding.

Hi! Before jumping into GPU sharding (which sounds a bit overkill on a 2GPU setup tbh) i'd encourage you to just reduce the batch size and increase the gradient accumulation...

Encountered a TypeError: expected string or bytes-like object when training the model using dual-GPU model sharding.

@ycemsubakan any chance that you could have a look to this?

Inconsistent Device Error in Explicit ASR Inference

@cyberso could you add ctc_lin: !ref to the modules list in the hyperparams yaml of the pretrained model and try again?

Inconsistent Device Error in Explicit ASR Inference

Alright then I believe it is an issue for @Adel-Moumen

noisy_wav, lens = batch.noisy_sig?

Hi, in this case the SB default sampler returns a tuple (batch, lengths of the sig in the batch) which is I believe, what is happening here.

Real-Time ASR with Device Capture using torchaudio StreamReader

@pplantinga @Adel-Moumen, this could be our first pipeline example, what do you think? I'm wondering if we could put some sore of voice activity detection in top of this to...