halo-livestream

Phrase-level segments from diarized Taglish (Tagalog–English code-switched) livestream recordings, processed for ASR and TTS training. Part of the halohalo Philippine-language dataset family.

Two configs:

Config	Sample rate	Contents
data/asr	16kHz mono	All gated segments, Whisper-ready
data/tts	24kHz mono	Strict subset: forced-aligned, non-overlapping, quality-gated

Usage

ASR / Whisper fine-tuning:

from datasets import load_dataset, Audio
ds = load_dataset("sapinsapin/halo-livestream", data_dir="data/asr")
# precise-timing subset:
ds = ds.filter(lambda x: x["alignment"] == "forced" and x["asr_cer"] <= 0.25)

TTS:

ds = load_dataset("sapinsapin/halo-livestream", data_dir="data/tts")
# per-speaker corpora via the namespaced speaker_id column

Schema

Field	Type	Description
audio	Audio	Mono WAV segment (16kHz asr / 24kHz tts)
sentence	str	Human transcription (bracket tags stripped)
language	str	tgl-eng (Taglish) or ISO 639-3 code
duration	float	Segment duration (s)
speaker_id	str	{recording_id}#S{n} — unique across recordings
gender	str	From source speaker profile
role	str	Speaker role description from source metadata
speech_type	str	spontaneous
source	str	Source recording id
start, end	float	Position in the source recording (s)
alignment	str	forced (MMS CTC) / interpolated / exact
align_score	float	Forced-alignment confidence (0–1), null if not forced
asr_cer	float	CER between transcript and Whisper large-v3 round-trip
overlap	bool	Heuristic overlapped-speech flag
speech_ratio	float	Fraction of segment covered by VAD speech
snr_db	float	Speech/nonspeech energy ratio within the source block
lufs	float	Integrated loudness
clip_ratio	float	Fraction of clipped samples

Processing

Block-level diarized transcripts were segmented per speaker turn, then upgraded with MMS-300M CTC forced alignment (romanization-based — handles code-switching), silero-VAD boundary snapping, and a faster-whisper large-v3 round-trip CER score per segment. See the pipeline documentation for stage details and export gate thresholds.

Splits

File-level split (md5(recording_id) % 10): speakers never cross splits. test populates as the corpus grows.

Quality caveats

• Spontaneous multi-party livestream audio: background music, notification sounds, and overlapping speech occur. Overlap is flagged heuristically, not removed from the ASR config.
• Transcripts are human-quality but unverified; asr_cer is provided as a per-segment quality signal — filter to taste.
• Speaker attribution errors in the source cannot be fully detected automatically (right words, wrong speaker label).
• Source audio is lossy 44.1kHz AAC; the 24kHz TTS config is suitable for training but is not studio-grade.
• Speech is from public livestream broadcasts and may contain real names and handles. Contact the maintainer for takedown requests.

Licensing

Research use. Refer to the source platform's terms for the underlying broadcast content.