Scicom-intl/Clean-Podcast

Clean Podcast & Movie Teacher Subsets

Clean 48 kHz / 16-bit / mono speech chunks used as teacher targets for the call-centre speech-restoration finetune (Malaysian/Singaporean telephony domain). These are the clean references only — the model learns to reconstruct this clean speech from a telephony-degraded version that is synthesised on the fly at training time (band-limiting, GSM/G.711-µ-law/MP3 codecs, line noise, VoIP dropouts). No degraded audio is stored here.

Each subset was filtered to keep only genuinely clean, single-speaker speech using DNSMOS P.835 bak (background-noise MOS) with a strict threshold of ≥ 3.644 — music, noisy, and overlapping-speech chunks are dropped.

Contents

Subset (zip prefix)	Chunks	Approx. clean audio	Source
podcast_sg_*.zip	14,193	~59.1 h	Singaporean podcast (malaysia-ai/singaporean-podcast-youtube)
podcast_my_*.zip	9,034	~37.6 h	Malaysian podcast (malaysia-ai/malaysian-podcast-youtube)
movie_my_*.zip	92	~0.4 h	Malaysian movie (malaysia-ai/malaysian-movie-youtube)
Total	23,319	~97 h	~27.4 GB across 9 zip parts

Each *.zip is a flat archive of .wav files (no internal directories; arcname = basename), split into ZIP_DEFLATED parts of ≤ 5 GB for upload. WAV filenames encode the source video title and YouTube id, e.g. <title> [<videoId>]_<chunk>.wav.

Audio format: PCM signed 16-bit, 48 000 Hz, mono, 15 s non-overlapping chunks.

Languages: predominantly Malay and English (incl. Singlish/Manglish), with Mandarin and Tamil present.

How it was built

For each source repo: download the archive (HF xet), selectively extract audio up to a duration budget, then per file: ffmpeg-decode → 48 kHz mono → 15 s non-overlapping chunks → score each chunk with DNSMOS P.835 → keep bak ≥ 3.644 as 48 kHz PCM_16 wav. Clean chunks are zipped (flat, ≤ 5 GB parts) and uploaded here. (Builder: prepare_podcast_clean.py in the Sidon call-centre pipeline.)

Usage

Download with the fast xet backend, then unzip (parts are independent):

import glob, os, zipfile
os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"
from huggingface_hub import snapshot_download

# grab one subset (or use allow_patterns=["*.zip"] for everything)
d = snapshot_download(
    "Scicom-intl/sidon-callcentre-podcast", repo_type="dataset",
    allow_patterns=["podcast_sg_*.zip"],
)
os.makedirs("podcast_sg", exist_ok=True)
for z in sorted(glob.glob(f"{d}/podcast_sg_*.zip")):
    with zipfile.ZipFile(z) as zf:
        zf.extractall("podcast_sg")   # flat *.wav land here

For parallel (distributed) extraction of all parts, see fetch_podcast_clean.py in the Sidon call-centre pipeline (xet download + multiprocessing unzip).

Provenance & intended use

Derived from publicly available YouTube audio (via the malaysia-ai/*-youtube collections), segmented and DNSMOS-filtered for research use as clean speech-restoration teachers. No speaker labels or transcripts are included. If you are a rights holder and want content removed, please open a discussion on this repository.

transcription subset — STT transcriptions

Whisper (whisper-large-v3-turbo) transcriptions of every chunk in the zips above, so the audio can be used where a transcript is required (voice-conversion reference prompts, ASR finetuning, filtering).

from datasets import load_dataset
ds = load_dataset("Scicom-intl/Clean-Podcast", "transcription", split="train")

23,274 rows, 96.97 h, 679 distinct source videos. Text and metadata only — the audio stays in the zips; join on zip + filename.

column	description
zip, filename	locates the chunk inside the archives above
video_id	YouTube id of the source video (the only speaker proxy)
subset	podcast_my, podcast_sg or movie_my
duration, sample_rate, channels	15.0 s, 48000 Hz, mono
transcription	STT output
char_len	transcription length in characters
sec_per_char	duration / char_len — see below
chars_per_sec	inverse, easier to read
looks_degenerate	ASR loop or near-empty output
stt_model	model that produced the transcription

Filtering

sec_per_char separates real speech from chunks whose transcript does not account for their audio. Median is 0.063 (~16 chars/sec); silence, music and missed speech land an order of magnitude higher ("OK." on 15 s scores 5.0). 89 rows are flagged looks_degenerate and 50 exceed 0.3 s/char.

clean = ds.filter(lambda r: not r["looks_degenerate"] and r["sec_per_char"] <= 0.3)

leaves 23,150 rows (99.5%).

Note the ratio is not language-neutral — Chinese packs more meaning per character than Malay or English, so split by script before tuning the threshold tighter.

Caveats

• Transcriptions are machine-generated and unverified — no human correction.
• Some chunks are singing, not speech. DNSMOS bak filtering removes noise, but sung vocals score as clean speech.
• movie_my is Turkish and Hindi film audio (with Malay subtitles), not Malay speech. The label describes the subtitle track, not what is spoken.