gijs/sea_speech

SEA Code-Switching

101,975 verified code-switched speech clips · 473.74 hours · 42,321 distinct source videos, covering Chinese, Thai, English, Hindi and Vietnamese.

Each clip contains at least one language switch by the same speaker, with the switch located in time.

By language pair

pair	clips	hours
en-hi	38,263	143.58
en-zh	37,964	207.57
en-th	15,035	75.45
en-vi	10,306	45.3
vi-zh	136	0.54
en-vi-zh	106	0.52
en-th-zh	73	0.39
th-zh	50	0.24
th-vi	12	0.03
en-hi-zh	12	0.06
en-th-vi	10	0.03
hi-zh	3	0.01
hi-th	3	0.01
en-hi-th	2	0.0

By switch level

value	clips
phrase	39,903
alternation	36,939
insertion	25,133

• tag — one embedded token · insertion — 2–4 · phrase — ≥ 5 tokens or ≥ 12 characters · alternation — ≥ 3 switches and ≥ 5 embedded tokens.

How a clip was verified

Transcribed with MOSS-Transcribe-Diarize 0.9B (timestamps + speaker labels), then:

• Script mixing decides Chinese / Thai / Hindi: a Han, Thai or Devanagari run adjacent to a Latin run is a switch.
• English vs Vietnamese share the Latin script, so tokens are resolved with wordfreq: Vietnamese diacritics first, then a frequency margin. Vietnamese is detected from the transcript, not the corpus label, because clips labelled yue/cmn that are really Vietnamese otherwise route down the wrong branch and every undiacriticked syllable (nay, ta, chi, ai) gets called English. A lone English token in a Vietnamese clip must beat Vietnamese by ≥ 0.8 zipf — may (+0.50) and top (+0.62) are Vietnamese in context, deadline (+4.12) and meeting (+5.15) are real insertions.
• Switches must stay within one speaker. Two speakers each using their own language is not code-switching; MOSS's diarization is what makes that separable.
• ASR boilerplate is stripped before counting ("Thanks for watching", "字幕由Amara.org社群提供"), which otherwise manufactures a switch on monolingual audio.
• Lone rare Latin tokens are dropped — "Aiden" in a Vietnamese sentence is a name. Ordinal and contraction fragments (th, st, re, ve) are blacklisted: ASR emits them standalone from "1st April", and zipf("th","en") = 4.20 is high enough to fool a plain threshold.

Fields

audio (16 kHz mono FLAC) · text · languages · language_pair · matrix_language · embedded_languages · switch_level · n_switches · spans (per-language text + start/end) · switch_points · duration · the original corpus LID label · full provenance (youtube_id, start_sec, end_sec, source_tar, source_key).

Usage

from datasets import load_dataset, Audio
ds = load_dataset("parquet", data_files="data/train-*.parquet", split="train")
zh_en = ds.filter(lambda r: r["language_pair"] == "en-zh")

# `datasets` >= 4 decodes Audio through torchcodec; without it, read the bytes:
import io, soundfile as sf
raw = ds.cast_column("audio", Audio(decode=False))[0]["audio"]["bytes"]
y, sr = sf.read(io.BytesIO(raw), dtype="float32")   # 16 kHz mono

Caveats

• Transcripts are machine-generated and unverified against human references.
• The selection = ml_l3 stratum is genre-biased toward language-lesson vocabulary drills.
• No English-matrix stratum was mined, so inter-sentential English-first switching is under-represented.
• Eval-quarantine flags were not joined; see the candidate card.