Desh Raj

In its current form, DOVER-Lap does not elegantly handle the case when the input hypotheses are mixed-type. By "mixed type", we mean that some of them have overlapping segments and...

Object of type 'LazyFilter' has no len()

5

I'm running into this weird issue, and I'm not sure why adding manifests should create a lazy iterator? ```python In [5]: m = prepare_ali_meeting('/export/c01/corpora6/AliMeeting') Preparing Train: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 209/209 [02:56 0...

Mismatch in number of samples between channels of `Recording`

2

Currently, when creating a multi-channel`Recording` manually by specifying sources as different single-channel files (as is done in the AMI setup, for example), we do not strictly enforce the condition that...

The following works well: ``` lhotse split -s 16 data/librispeech/raw_cuts_dev-clean.jsonl.gz data/librispeech/tmp ``` But the following does not: ``` gunzip -c data/librispeech/raw_cuts_dev-clean.jsonl.gz | lhotse split -s 16 - data/librispeech/tmp ``` Here...

Check failed: axis < NumAxes() in ragged.cu

2

I trained a Conformer CTC model with Icefall and was trying to use it for decoding, but I am getting the following error: **[F] /export/c07/draj/mini_scale_2022/k2/k2/csrc/ragged.cu:116:k2::Array1& k2::RaggedShape::RowIds(int32_t) Check failed: axis <...

Standard recipe for fine-tuning

1

Is there a recommended recipe I can refer to for fine-tuning one of the pretrained models (say, the Gigaspeech model) on my own data?

VAD workflow with Silero

2

In terms of which VAD to apply, you can use e.g. SileroVAD: https://github.com/snakers4/silero-vad/wiki/Examples-and-Dependencies#examples Actually a workflow/integration into Lhotse would be nice if somebody is willing to contribute that. _Originally posted...

Use torchaudio to read sph files first

3

New torchaudio with ffmpeg backend can read SPH files. It doesn't seem to be any faster than sph2pipe, but at least it does not require installing new tools.

[Not for merge] Diarization workflow with SpeechBrain

4

This workflow shows how we can use SpeechBrain x-vectors + sklearn agglomerative clustering to perform a crude speaker diarization. This can be used on top of the whisper workflow to...

It seems that currently, we never enforce `hop` to be positive. I am not sure what the intended result would be for zero or negative hop values.