How does MDX_CHUNK work?

[FlippFuzz]

Flag that I'm referring to:

  --mdx_chunk MDX_CHUNK
                        Chunk size in seconds for MDX-Net filter. Notes:
                        Smaller is using less memory, but can be slower and
                        produce a bit lower quality. (default: 15)

---

Command:

I'm dealing with a song, some parts are cut off, so I'm messing with lowering vad_threshold.

Faster-Whisper-XXL/faster-whisper-xxl --model large-v2 --ff_mdx_kim2 --mdx_chunk 64 --language ja --vad_method pyannote_v3 --vad_threshold 0.3 --task transcribe --output_format srt -- "/kaggle/input/video.mp4"

---

Some tests

There don't seem to be much differences.
I can't really tell whether the quality is different too.

mdx_chunk 15

GPU RAM - 4.3 GiB/15 GiB

MDX Kim_Vocal_2 model probably is running on CUDA:   6% | 41/661 | 00:59<<10:28

mdx_chunk 32

GPU RAM - 6.4 GiB/15 GiB

MDX Kim_Vocal_2 model probably is running on CUDA:   9% | 27/310 | 01:19<<09:19 

mdx_chunk 64

GPU RAM - 13.5 GiB/15 GiB

MDX Kim_Vocal_2 model probably is running on CUDA:   8% | 12/155 | 01:37<<09:50