Piotr Żelasko

Just throwing the idea around in case somebody wants to pick it up, I wanted to do it for some time but can't find the time. There's a new position...

You are one of the main groups that was using Lhotse throughout the last year. I’d like to solicit your feedback on Lhotse’s API — do you think there is...

Building lexicons in Python

16

The current setup inherits building lexicon FSTs from Kaldi. I think it makes sense to have the ability to build it directly in Python, which should make building new recipes...

Inconsistency between the reported loss and the actual loss used for gradient computation

2

The loss we backprop is normalized by the number of supervisions: https://github.com/k2-fsa/snowfall/blob/5d1b00dd2ab4c809714f588fa6d2487cde5ea46c/egs/librispeech/asr/simple_v1/mmi_att_transformer_train.py#L114-L117 But the loss we report is normalized by the number of frames: https://github.com/k2-fsa/snowfall/blob/5d1b00dd2ab4c809714f588fa6d2487cde5ea46c/egs/librispeech/asr/simple_v1/mmi_att_transformer_train.py#L267-L278 It looks like it wasn't...

[WIP] 2-state HMM topo as an alternative to CTC topo

8

I'm trying to build a topology where the "blank" is phone-specific instead of shared between phones (I believe that corresponds to Kaldi's chain topology). I added a function `build_hmm_topo_2state` which...

Random sampling with concatenation vs bucketing

21

FYI I ran the transformer training with mmi + attention with different batch sampling settings to check which sampler gives better WER. The results are from the averaged model using...

"Check failed" after decoding + shortest_path

8

I've been playing around with the recipe to get to know k2 better. I trained the model for 10 epochs and then tried to perform decoding with the following code...

New `ffmpeg` backend changes samples when saving WAVE

2

### 🐛 Describe the bug Snippet to reproduce the error is provided below. Adding `backend="sox"` or `backend="soundfile"` to `torchaudio.save` removes the issue. ```python import os from tempfile import NamedTemporaryFile os.environ["TORCHAUDIO_USE_BACKEND_DISPATCHER"]...

In my personal experience it's rarely useful to have a separate RecordingSet and SupervisionSet, and most of the time the users seem to build a CutSet out of them right...

Setting this up so I remember to clean this up. One can get an infinite CutSet by calling `cuts.repeat()`, but generally when working with multiple cutsets that represent very large...