sapinsapin/kumu-livestream-segmented

halo-livestream

Real Taglish code-switching from livestreams — every segment carries forced-alignment confidence, ASR round-trip CER, SNR, loudness and overlap flags.

62 segments · 3 speakers · seed release

🌱 This is a seed release — 62 segments, about 7 minutes

It exists to publish the pipeline and the schema, not to be a training corpus. Nothing here is big enough to train on. What is worth your time is the per-segment quality metadata below — and the processing code, which scales to as many recordings as you feed it.

If you want volume today, use sapinsapin/pld (448 h) or sapinsapin/filipinospeechcorpus (65 h).

Why it exists: Filipino speakers switch between Tagalog and English mid-clause, constantly. Studio corpora don't capture it because prompts are written in one language. This pipeline targets natural code-switched speech and attaches enough quality signal per segment that you can set your own bar instead of trusting an opaque "clean" label.

---

Quickstart

from datasets import load_dataset

asr = load_dataset("sapinsapin/halo-livestream", "asr", split="train")
tts = load_dataset("sapinsapin/halo-livestream", "tts", split="train")

print(asr[0]["sentence"])
# 'Yes mi. Saglit lang, saglit lang pa- paalis paalis- paalis ako eh.'

Filter to precisely-timed, cleanly-transcribed segments — the pattern this dataset is really shipping:

good = asr.filter(lambda x:
    x["alignment"] == "forced"      # MMS CTC alignment, not interpolated
    and x["asr_cer"] <= 0.25        # transcript agrees with Whisper round-trip
    and not x["overlap"]            # no overlapping speech
    and x["speech_ratio"] >= 0.6    # mostly speech, not silence
)

---

The two configs

Config	Rate	Segments	Audio	Purpose
asr	16 kHz	53	~6.4 min	All gated segments, Whisper-ready
tts	24 kHz	9	~1 min	Strict subset: forced-aligned, non-overlapping, tighter quality gates

The tts config is deliberately small — it is what survives TTS-grade gating, and the ratio (9 of 53) is itself the useful signal about how much livestream audio is actually usable for synthesis.

	asr	tts
Forced-aligned	49 / 53	9 / 9
Median round-trip CER	0.203	0.087
Mean segment length	7.3 s	6.3 s

---

Schema

Standard fields:

Field	Type	Description
audio	Audio	Mono segment (16 kHz asr / 24 kHz tts)
sentence	str	Human transcription, bracket tags stripped
language	str	tgl-eng (Taglish) or ISO 639-3
duration	float	Segment length (s)
speaker_id	str	{recording_id}#S{n} — unique across recordings
gender, role	str	From the source speaker profile
speech_type	str	spontaneous
source	str	Source recording id
start, end	float	Position within the source recording (s)

Quality metadata — the part that makes this dataset useful. Every segment is scored, so you can pick a threshold instead of accepting someone else's:

Field	Type	Description
alignment	str	forced (MMS-300M CTC) / interpolated / exact
align_score	float	Forced-alignment confidence 0–1; null when not forced
asr_cer	float	CER between the human transcript and a faster-whisper large-v3 round trip — the single best "is this transcript right" signal
overlap	bool	Heuristic overlapping-speech flag
speech_ratio	float	Fraction of the segment covered by VAD speech
snr_db	float	Speech/non-speech energy ratio
lufs	float	Integrated loudness
clip_ratio	float	Fraction of clipped samples

---

How it was built

1. Diarized transcripts at block level, segmented per speaker turn.
2. Forced alignment with MMS-300M CTC, romanization-based — this is the part that survives code-switching, where a Tagalog-only or English-only aligner drifts at every switch point.
3. Boundary snapping with silero-VAD, so segments start and end on speech.
4. Round-trip scoring: transcribe each segment with faster-whisper large-v3 and record CER against the human transcript.
5. Gating into the asr and tts configs by alignment quality, overlap, CER, and loudness/clipping thresholds.

Pipeline source: process_livestream.py · docs: docs/livestream_pipeline.md

---

Models trained on this data

None — 62 segments is far too little, and publishing a model trained on it would be misleading.

The trainers accept this dataset with a flag (--dataset livestream), so once the corpus grows the recipe is already wired:

python finetune_asr.py --dataset livestream --push   # → whisper-small-halohaloLS
python finetune_tts.py --dataset livestream --push   # → speecht5_tts-halohaloLS

Working baselines on the sibling Filipino corpus, for reference: speecht5_tts-fsc · whisper-small-fsc

---

Limitations

• Tiny. 62 segments from a single source recording and 3 speakers. Any metric computed on it is noise.
• Transcripts are human but imperfect — median round-trip CER is 0.203 on the asr config, which reflects both genuine transcription variance and the fact that Taglish orthography is unstandardised (nag-aano / nagaano / nag aano).
• CER is the honest metric here, not WER: word-level scoring punishes legitimate spelling variation in code-switched text.
• Numerals are not verbalized — digits appear as digits in transcripts.
• Speaker roles come from source metadata and are not independently verified.
• Livestream audio carries background music, stream artifacts, and variable mic quality; snr_db and clip_ratio are there so you can see it.

---

Related datasets

Part of the halohalo Philippine-language speech family:

Dataset	What it covers	Scale
halo-livestream (this one)	Taglish code-switched livestream speech	62 segments · seed
sapinsapin/pld	10 Philippine languages, prompted	334k utterances · 448 h
sapinsapin/filipinospeechcorpus	Filipino studio read + spontaneous	305k segments · 65 h

---

Terms and ethics

No license is asserted over the source audio. It is livestream content whose rights belong to the original broadcasters; this repo publishes derived segments and metadata for research use only. Check the source terms before redistributing, and treat this as research material, not a licensed corpus.

Speakers are identified only by an opaque {recording_id}#S{n} key, with no names, handles, or channel identifiers in the data. If you are a speaker in this data and want a segment removed, open a discussion on this repo and it will be taken down.

Do not use this data to identify, profile, or synthesize the voice of any individual speaker without their consent.

---

Contributing

The pipeline is the point — it scales to whatever you feed it, and every stage is documented.

• Run it on your own recordings: github.com/sapinsapin/halohalo
• Report bad segments via the Community tab (include source and start)
• Taglish/code-switching evaluation sets are badly needed; contributions welcome