Datasets:

---

VITA-MLLM

/

VITA-Audio-Data

like 7

Follow

VITA-MLLM 69

Hi, thanks for releasing this dataset!

I noticed that in the provided JSONL files, the audio field points to .wav paths like:

datasets/MLCommons/peoples_speech/wav/41/418bf9367a72...

But as far as I know, People’s Speech is stored in .parquet format, not pre-extracted .wav.
Converting all .parquet files into .wav would take a huge amount of disk space and time.

So I’d like to ask:

• Did you actually convert the entire dataset into .wav files to create the JSONL?
• Or is there a script that maps .parquet audio data directly to these .wav paths without full extraction?
• And if possible, could you share the code you used to construct the JSONL dataset?

Any clarification would be super helpful. Thanks!