BS-RoFormer, 178 Hours, and One Unverifiable Link: A Dataset Claim Under the Microscope

𝟭𝟳𝟴 𝗵𝗼𝘂𝗿𝘀 𝗼𝗳 𝗰𝗹𝗲𝗮𝗻 𝗠𝗮𝗹𝗮𝘆𝗮𝗹𝗮𝗺 𝘀𝗽𝗲𝗲𝗰𝗵. 𝗙𝗿𝗲𝗲. 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗹𝗼𝘄-𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀 𝘀𝘁𝗮𝘆 𝗹𝗼𝘄-𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲 𝘂𝗻𝘁𝗶𝗹 𝘀𝗼𝗺𝗲𝗼𝗻𝗲 𝗼𝗽𝗲𝗻𝘀 𝘁𝗵𝗲 𝗱𝗮𝘁𝗮.
Malayalam has ~38 million speakers and almost no open speech corpora. That's the whole reason its voice models lag behind.
So here's Malayalam-Speech-178h:
🎙️ 86,799 clips, 48 kHz
🧹 Fully cleaned — vocal isolation (BS-RoFormer) → de-reverb → denoise → enhancement
⏱️ 178 hours of processed speech
🆓 Public on Hugging Face
One thing up front: this is unlabeled audio. No transcripts. It's built for self-supervised pretraining, voice/speaker modelling, or as raw material for your own labelling pipeline — not for supervised ASR out of the box.
I'd rather tell you that now than have you download 123 GB and find out.
The expensive part is already done. Source separation and denoising at this scale takes serious GPU time — that's what you're getting for free.
from datasets import load_dataset
ds = load_dataset("psk/malayalam-speech-178h", split="train")
🔗 Link in the first comment.
If you're working on Malayalam speech — or any Indian language — take it and build something.
And if you're sitting on processed audio nobody else has, consider putting it out there. The gap in Indian-language speech data doesn't close by itself.

I checked what I could verify. Here's the breakdown:

Could not verify — the core claim:

• I could not find "psk/malayalam-speech-178h" on Hugging Face through search, and I can't fetch a Hugging Face URL directly unless it's already surfaced in search results. This doesn't prove it's fake — it could simply be too new to be indexed, or search coverage of individual HF dataset pages is patchy — but I can't confirm the dataset exists, nor verify its specific stats (86,799 clips, 178 hours, 48kHz, 123GB) independently. If you have the direct link, share it and I can check it properly.

Technically plausible / accurate framing:

• BS-RoFormer is a real, well-regarded architecture for music/vocal source separation (band-split RoFormer) — commonly used for vocal isolation. Using it as a first step in a speech-cleaning pipeline (isolate vocals → de-reverb → denoise → enhance) is a legitimate, realistic workflow, not invented jargon.
• The claim that this is "unlabeled audio... not for supervised ASR" is an internally consistent and honest caveat — a real, responsible detail for a raw self-supervised pretraining corpus.

Overstated or questionable:

• "Almost no open speech corpora" for Malayalam is an exaggeration. There are several existing open Malayalam speech resources: OpenSLR's Malayalam corpus, AI4Bharat's IndicVoices (23.7K hours across 22 Indian languages including Malayalam), IIT Madras's IndicTTS Malayalam corpus, ICFOSS's IMaSC corpus, and multiple Whisper fine-tuning corpora on Hugging Face. Malayalam is genuinely under-resourced relative to English/Hindi, but "almost no" open corpora isn't accurate — "far less than it needs" would be fairer.
• "~38 million speakers" — plausible but on the high end. The 2011 Indian census puts native Malayalam speakers around 34–35 million; ~38 million figures typically include diaspora/second-language speakers. Not wrong, just not the most commonly cited number.

Bottom line: The technical pipeline described (BS-RoFormer → de-reverb → denoise) is credible and the "unlabeled, not ASR-ready" disclaimer is a good transparency signal, but I can't independently confirm the dataset itself exists at that exact HF path with those exact stats, and the "almost no open corpora" framing overstates the actual scarcity. If you can share the actual link, I can verify the file counts/hours/license directly.