These are unsegmented livestream recordings of identifiable people speaking conversationally. Access is granted for speech research and dataset construction. By requesting access you agree not to redistribute the recordings, not to attempt to identify or contact the speakers, and not to use the audio for voice cloning or impersonation of the speakers.

halo-livestream-raw

Unsegmented source recordings behind sapinsapin/halo-livestream — the archival input to the pipeline, not a training set.

1 recording(s) · 26:56 · 9.0MB

🔒 Gated on purpose

Full-length conversation between named, identifiable speakers is a very different privacy proposition from the short segments in the processed dataset. Access is manually reviewed. If you want segmented, quality-scored audio for training, use sapinsapin/halo-livestream instead — it needs no approval.

What this is

sapinsapin/halo-livestream publishes ~30-second segments with forced-alignment confidence, round-trip CER, SNR and overlap flags. Those segments are derived. This repository holds what they were derived from, so results stay reproducible and future pipeline versions can be re-run over the same input without re-collecting anything.

audio/{file_id}.{m4a|mp3|opus|flac}   archival audio track
transcripts/{file_id}.json              operator transcript, as delivered
index.jsonl                               one row per recording

Recording	Duration	Audio	Size	Speakers
`9e053c5b6f4b2a7b…`	26:56	aac 48000 Hz / 1 ch	9.0MB	3

Languages present: Taglish · speaker entries: 3

How the audio is stored

Audio is kept as close to as-delivered as possible:

Already-compressed sources (AAC/MP3/Opus) are stream-copied, never re-encoded. Transcoding lossy audio to FLAC cannot recover what the encoder discarded, and it inflates size roughly tenfold — on the seed recording, 9.0 MB AAC becomes 113 MB FLAC with a different decoded checksum. The copied stream round-trips bit-exactly.
Only uncompressed PCM is encoded, to FLAC, where compression is lossless and also sidesteps the 4 GB WAV ceiling that multi-hour streams hit.
Video tracks are dropped. The pipeline never reads them, and they carry the most personal data.

index.jsonl records audio_sha256 for every track so integrity is checkable after download.

Transcript schema

{
  "metadata": {
    "file_properties":  { "duration": "00:26:55", "audio_specifications": {...} },
    "linguistic_profile": { "primary_language": "Taglish", "content_theme": "..." },
    "speaker_profile":  { "speaker_count": 3, "speakers": [{"speaker_id": "Speaker 1", ...}] }
  },
  "transcription": [
    { "time_range": "02:25 - 02:33",
       "dialogue": [{"s": "Speaker 1", "txt": "..."}] }
  ]
}

Timings are block-level only — a block can bundle several turns with no per-turn timestamps. The pipeline's parse stage interpolates a baseline from character counts and the align stage replaces it with forced alignment. Treat the time ranges here as approximate.

Reported file_properties come from the operator and may not match the container; trust index.jsonl, which is probed from the actual file.

Rebuilding the processed dataset

huggingface-cli download sapinsapin/halo-livestream-raw --repo-type dataset --local-dir raw/
export LIVESTREAM_DIR=raw/
python process_livestream.py --stages parse,align,qc,export --push

Discovery accepts this layout directly, so a downloaded snapshot feeds straight back into the pipeline.

Limitations

Small. This is a seed archive, not a corpus.
Diarization and transcription are operator-supplied, not verified here.
Speaker labels are per-recording; Speaker 1 in two files is not the same person. The pipeline namespaces them at parse time.

License and consent

Recordings were collected from public livestreams. No consent was obtained from speakers for model training specifically — treat this as a research artifact and honour the gated terms. If you are a speaker in one of these recordings and want it removed, open a discussion on this repository.

Downloads last month: 7

Total file size:

9.5 MB

You need to agree to share your contact information to access this dataset