trunglebka

The above error likely due to supervision start is relative to cut start, but if I set `cut.supervisions[0].start = 0` the old error occurred: ``` ValueError: Cannot load features for...

@pzelasko I believe it's the correct cutset, but I'm not entirely sure because I'm not proficient enough with Lhotse. Regarding to the error `AssertionError: Supervisions ending after the cut are...

@desh2608 The dataset was be prepared with default [trim_to_supervisions](https://gist.github.com/trunglebka/725ec0a8257b9369ac9f9a52d9a862b3#file-preprocess_nsc-py-L136) (keep_overlapping=True copied from gigaspeech recipe) Preparing process take time and I encountered many problems with nsc dataset so in icefall's training...

Regarding to processing time of compute_and_store_features_batch, I think that it mainly because of disk IO. So we can create special kind of SimpleCutSampler that prefetch audio into memory using a...

Could you please elaborate on the problem, our dataset has many quote `"` and I'm not sure what is the problem with it?

I mean why do quotes cause Cuda OOM, is there parsing problem with training code but not preprocessing?...

> hoverZ You made my day. But can you elaborate about the problem?

Just an additional information. I'm using python API of triton. For the same model, the first run of "model.infer" take 5s but "model.async_infer" take 36s. Second run is almost the...

Mine just generate "they are sleeping..." and it cost $5 :smiley:

@singhniraj08 I'm facing similar problem as the OP, since the `finish_reason` is always `OTHER` and the issue you linked is talking about safety I'm not sure it is related problem....