twsyllables — Taiwanese Mandarin syllable acoustics

Per-syllable acoustic reference data for Taiwanese Mandarin* (臺灣華語, cmn-Hant-TW): 37,947 measured syllable tokens, position-sensitive acoustic templates for 1,491 syllable×tone types, voice-onset-time norms for all 17 obstruent initials, and a between-speaker variability model estimated over 271 speakers. Every number was measured from native Taiwanese recordings by one reproducible pipeline; no figure in this dataset is hand-set or copied from the literature.

To our knowledge this is the first openly licensed set of syllable-level acoustic norms for adult Taiwanese Mandarin. The existing spoken corpora — the Sinica Taiwan Mandarin Conversational Corpus (43 hours, 170 speakers, orthographic transcription) and the NCCU Corpus of Spoken Taiwan Mandarin (conversations in discourse-analytic transcription) — distribute recordings and transcripts for discourse research, not acoustic measurements. Descriptive phonetic values for Taiwanese Mandarin otherwise live scattered across individual papers, usually a handful of speakers per study, and Beijing-normed references do not transfer: Taiwanese Mandarin realizes tone 3 as a low tone without the final rise, keeps unaspirated stops with short but positive VOT, and merges or weakens the retroflex series in ways a mainland norm would penalize incorrectly.

0.4.0 measures the medial and the nucleus through separate windows. In 0.3.0 one window covered the whole rime and sampled the glide wherever a medial opened it, so f1_vowel read below f1_mid — impossible for a nucleus. Splitting the window closes that gap (-ia −65 → −2 Hz, -ie −54 → −16, -ian −26 → −13) and improves discrimination on held-out speakers: final-axis AUC 0.5420 → 0.5586, initial-axis AUC 0.7221 → 0.7303, aspiration and sibilant-place decisions +0.6 pp each, native tokens scored green 52.8% → 53.4%, top-1 unchanged. tone_contour.n_band_speakers is new; no field was removed; f1_vowel, f2_vowel, f1_onset, f2_onset and the ratios built on them change value. See Changelog → 0.4.0.

0.3.0 measures the vowel where the vowel is. The rime was read at the midpoint of the voiced span, which for a nasal coda lands in the murmur: -ang came out with a lower first formant than -a, which no vocal tract does. The vowel is now read over the steady part of the rime and normalized against the speaker's own pitch. Eleven token fields, six template fields, four tone_contour subfields and the apical rime flag are new; nothing was removed. See Changelog → 0.3.0.

0.2.0 corrected data errors in 0.1.0. 234 of 42,009 token measurements (0.56%) were filed under the wrong syllable key, and nine keys carried a bopomofo spelling that contradicted the key. Both are fixed; see Changelog → 0.2.0 → Corrections before reusing 0.1.0.

Status: active development. These numbers improve continuously — the measurement pipeline, the templates and the calibration are re-derived as the underlying corpora and the extraction code improve, and released versions will not be byte-stable between updates. Pin a release tag (revision="v0.2.0" in load_dataset, see Versions) for anything that must reproduce. The scoring engine that consumes this dataset in production will be released as an open Ruby gem, twspeech, with a matching extraction specification and conformance fixtures; a thin Python scorer is planned alongside it.

At a glance

config	rows	contents
syllables	1,491	the inventory: one row per syllable×tone key — zhuyin, pinyin, example characters, token counts, confusion sets
templates	6,400	acoustic reference templates: 1,491 citation + 1,491 word + 1,452 word-initial + 983 word-medial + 983 word-final
tokens	37,947	per-recording syllable measurements: ~40 acoustic features each, with provenance
variability	29	between-speaker dispersion per feature, estimated over 271 Common Voice speakers
vot_norms	17	pooled voice-onset-time distributions per initial
quality	965	leave-one-out recognizability of each syllable type against its confusion set

Plus calibration/ — the runtime calibration of the scoring system (axis norms, verdict thresholds, style factors, speaker pitch references and the contextual bending of tone and length in connected speech) as raw JSON, documented below.

Formats. The two large configs are Parquet (zstd, typed schema baked in): columnar reads mean pulling one feature out of 37,947 tokens costs under a megabyte, and DuckDB/Polars can query them over hf:// without downloading the file. The small configs stay as human-readable JSONL. MANIFEST.json carries row counts and SHA-256 of every file.

Source corpora

corpus	recordings	speakers	license	role here
MOE 國語辭典簡編本 audio (characters and words)	19,431 clips (including the narration row below)	1 institutional voice (moe_tw)	CC BY-ND 3.0 TW	token measurements, template centers
當代中文課程 (A Course in Contemporary Chinese) narration	— (counted above)	1 narrator (dangdai_tw)	proprietary; measured only, never redistributed	token measurements, word-position templates
全字庫 CNS 11643 syllable audio	2,933 clips	2 studio voices (cns_f, cns_m)	OGDL v1.0	token measurements, segmental template support
Mozilla Common Voice 26.0 zh-TW, speakers with a declared Taiwan birthplace	6,557 clips	274 pseudonymous speakers (271 usable for the model)	CC0	between-speaker variability widths only

No audio is included or redistributed in this dataset. What is published is acoustic measurements — numeric facts about the recordings — plus aggregates over them. Token counts per source: moe_word 19,720, moe_char 5,247, dangdai 10,046, cns_voice 2,924. See NOTICES.md for the authoritative licensing statement, including the required 全字庫 attribution and the legal basis on which measurements from ND-licensed audio are published.

How the numbers were measured

All audio is decoded to mono at 22,050 Hz and analyzed on a 10 ms hop by a pure-Ruby DSP stack (dsprb for FFT, MFCC, YIN f0 and LPC formants; dtwrb for dynamic time warping, DBA barycenter averaging and robust dispersion). The chain, in brief:

1. Segmentation. Speech extent by adaptive energy thresholds robust to room tone and digital silence; multi-syllable recordings are split at energy valleys, with a separate path for deliberately paused, syllable-by-syllable delivery.
2. Per-syllable features. f0 track (YIN) cleaned, octave-corrected and resampled to a 16-point contour in semitones around the token's reference f0; three LPC formant tracks, published resampled to 8 points but read for the vowel on the analysis frames themselves — the medial over frames 5–25% of the voiced span, the nucleus over 30–70%, or 25–60% where a nasal closes the rime and the later frames are murmur; 13 MFCC coefficients at 12 time points; spectral moments of onset frication; nasal-tail measures; durations. Voice onset time comes from a dual-envelope method (voiced energy below 800 Hz against burst energy above 1,200 Hz), is measured at any position in a word — utterance-initial, after a pause, or in the valley between syllables — and carries an explicit reliability flag (vot_reliable) that is true only when a release burst was genuinely located rather than assumed at the analysis-window edge.
3. Robust statistics. Every scalar feature is summarized as median, MAD, SD, p05/p95 and extremes; outliers are dropped before template construction and the drop count recorded. Tone contours are averaged by DBA barycenter with per-point dispersion.
4. Position-sensitive templates. Five template styles per syllable×tone key, built from citation forms, in-word occurrences, and word-initial, word-medial and word-final occurrences respectively, with documented blending when a position has too few tokens (the provenance.style field says exactly what was blended).
5. Variance decomposition. Within-speaker widths come from the token corpora; between-speaker widths are estimated per feature over 271 Common Voice speakers (25,029 syllable measurements) and folded into every template's sigma (sigma² = sigma_within² + sigma_between²), after a style correction measured between read and citation speech.
6. VOT pooling. Syllable-level VOT samples are pooled per initial (vot_norms) and templates for sparse keys draw on the pool — the pooled, n_key and n_pool fields keep that traceable.

Aspiration comes out textbook-clean at corpus scale — pooled VOT medians:

pair	plain	aspirated
ㄅ / ㄆ (b / p)	11 ms	88 ms
ㄉ / ㄊ (d / t)	12 ms	89 ms
ㄍ / ㄎ (g / k)	26 ms	93 ms
ㄐ / ㄑ (j / q)	71 ms	132 ms
ㄗ / ㄘ (z / c)	60 ms	127 ms
ㄓ / ㄔ (zh / ch)	59 ms	113 ms

Configs

Zhuyin is the primary spelling; identifiers are pinyin. Every config that names a syllable or an initial carries its bopomofo in a zhuyin column, complete for all 1,491 keys and ordered before pinyin. The machine identifier, however, stays ASCII — keys like kan4 (ㄎㄢˋ) survive filesystems, URLs, shells and Unicode normalization untouched, sort predictably, and parse trivially (key[-1] is the tone). Bopomofo would make a fragile identifier — tone 1 is unmarked, the other tone marks are spacing modifier letters, and macOS and Linux normalize such filenames differently — so it is data, not a key.

syllables

One row per syllable×tone key in the template inventory.

field	type	meaning
key	str	syllable + tone digit, e.g. bu4 (ㄅㄨˋ); tone 5 is the neutral tone
zhuyin	str	bopomofo with tone mark, e.g. ㄅㄨˋ; complete for every key. The tone mark is the citation tone, the key digit is the tone spoken. They differ for third tones raised to tone 2 by 3-3 sandhi: ban2 carries ㄅㄢˇ because the tokens are 板 bǎn produced as bán
syllable, tone, pinyin	str/int	segmental syllable, tone 1–5, pinyin key
chars	list[str] | null	example characters read with this key (null for 249 keys added after the character index was built)
n_tokens, n_citation	int	tokens used by the citation template; citation-form tokens among them
style	str	what the citation template was actually built from (citation, mixed, …)
confidence	str	high / medium / low / very_low, from token and speaker counts
sources	dict	token counts per source corpus
confusions	list[str] | null	the rival keys this syllable is tested against in quality
apical	str | null	dental for the empty rime 空韻 after ㄗㄘㄙ ([ɿ]), retroflex after ㄓㄔㄕㄖ ([ʅ]), null elsewhere. Pinyin writes all three as i; only the last is the close front vowel [i]

templates

One row per key×style; 6,400 rows across five styles (citation, word, word_initial, word_medial, word_final). Every scalar feature is a stat block:

{"median": ..., "mad": ..., "sd": ..., "p05": ..., "p95": ..., "min": ..., "max": ...,
 "n": ..., "sigma_within": ..., "sigma_between": ..., "sigma": ...}

sigma is the tolerance the scoring system actually uses; sigma_within and sigma_between are its decomposition (present where the variability model applies; vot_ms blocks add n_key, n_pool, pooled). Curve features carry center and per-point sigma arrays instead.

field group	fields	notes
identity	key, style, zhuyin, syllable, tone, pinyin, norm	norm is the norm family, currently always taiwan
structure	structure.{initial, initial_ipa, medial, nucleus, final, coda, nasal_coda, aspirated, sibilant, apical}	phonological parse; initial_ipa carries the Taiwan value, so ㄏ is [h] rather than [x]
provenance	provenance.{n_tokens, n_dropped_outliers, n_speakers, speakers, sources, n_isolated, n_citation, n_available, style, confidence}	exactly what the template was built from
tone	tone_contour (16-pt center, sigma, spread, low, high, semitones, plus n_band, n_band_speakers), tone_range, tone_slope, f0_register	contour is relative to the token-level reference f0. low/high are the band inside which the productions behind this template actually fall — the central 80% per point, widened by the between-speaker model; for most keys those productions come from a single speaker, so read the band as one native's range and not as the language's; spread is the per-point dispersion of the same unwarped curves, n_band the number of curves behind them, n_band_speakers how many speakers those curves come from — 788 of the 804 word-style bands rest on one — and sigma the older DTW-aligned one. f0_register places the reference in the speaker's range
duration	duration_ms, voiced_ms, voiced_ratio
onset	vot_ms, vot_ratio, fric_ms, fric_centroid, fric_spread, fric_skewness, fric_kurtosis	onset noise spectral moments in Hz where dimensional
vowel	f1/f2/f3 (8-pt tracks, Hz), f1_over_f0, f2_over_f1, f2_onset_ratio, f1_onset_over_f0	what the scorer reads: height over the speaker's median pitch and frontness, both from the nucleus window, against the same pair read from the medial window
vowel, 0.2	f1_mid, f2_mid, f3_mid, f1_ratio, f2_ratio, f2_end_ratio, f2_delta_ratio	the midpoint measurements of 0.1.0–0.2.0, unchanged and still published
coda	nasal_ratio_tail, nasal_ratio_mid, energy_tail_ratio, f2_end_over_f1, nasal_antiformant, centroid_ratio	nasality at the end against nasality in the middle, the energy fall into the coda, and the second-formant offset
spectral	mfcc (12×13 center+sigma), mfcc_scale	13 coefficients at 12 time points
bookkeeping	variability_applied	whether between-speaker widths were folded in

tokens

One row per measured syllable occurrence — the raw material behind the templates, published so that others can fit their own models. A handful of keys appear here with too few tokens to have earned a template, so joining tokens to templates leaves 25 keys unmatched by design.

field	type	meaning
key	str	syllable×tone key
clip	str	source recording filename (audio not included; joins across rows and, for OGDL/CC0 sources, with the upstream corpora)
syllable_index, n_syllables	int	position of this syllable in the recording and total syllables in it
speaker	str	moe_tw, dangdai_tw, cns_f, cns_m
source	str	moe_char, moe_word, dangdai, cns_voice
transcript	str	the word or character read
scalar features	float	durations, f0_ref_hz, tone range and slope, VOT (+vot_reliable), frication moments, formant measurements (+formants_reliable), nasal measures
f1_vowel, f2_vowel	float	formants over the nucleus: analysis frames 30–70% of the voiced span, or 25–60% where a nasal closes the rime
f1_onset, f2_onset	float	formants over the medial and the consonant-to-vowel transition: analysis frames 5–25% of the voiced span
f1_over_f0, f1_onset_over_f0	float	those divided by the speaker's median pitch from calibration/speaker_pitch.json
tone_curve	list[16]	pitch contour, semitones relative to f0_ref_hz
mfcc	list[12]×[13]	MFCC trajectory
f1, f2, f3	list[8]	formant tracks, Hz
energy_curve	list[16]	energy envelope, dB

variability

One row per feature: feature, mode (absolute — sigma in the feature's own units; relative — sigma as a fraction), sigma_between (scalar), and sigma_between_curve (16-point array, for tone_contour only). Estimated per feature over syllables attested by at least three distinct speakers in Common Voice 26.0 zh-TW, restricted to speakers who declared a Taiwan birthplace; 271 speakers, 25,029 syllable measurements. The template centers never come from this material — only the tolerance widths do.

vot_norms

One row per obstruent initial — ㄅ ㄆ ㄈ ㄉ ㄊ ㄍ ㄎ ㄏ ㄐ ㄑ ㄒ ㄓ ㄔ ㄕ ㄗ ㄘ ㄙ (b p f d t g k h j q x zh ch sh z c s); the sonorants ㄇ ㄋ ㄌ ㄖ (m n l r) carry no VOT. Columns: zhuyin, initial (pinyin), then the pooled VOT stat block across all syllables sharing that initial. This is the pool sparse templates draw from.

Why the pool is not split by the following vowel. It could be: the vowel effect is real and large in this corpus, reproducing the published finding that high vowels lengthen a preceding stop's VOT — ㄍ g runs 16 ms before -a against 31 ms before -u, ㄗ z 48 against 71 ms before -i. Splitting the pool that way was built and measured, and it moved the initial axis backwards (d′ 1.173 → 1.168, AUC 0.832 → 0.830). The reason is structural: 95.7% of the rival pairs that differ in their initial share the same following vowel, so conditioning on the vowel shifts a syllable and its confusable rivals by the same amount and buys no separation. A narrower pool also stops being a class — one bucket, ㄙ s before -e, was 72% a single syllable, and its median leaked into every other s+e template. The pool stays keyed on the initial alone.

quality

Leave-one-out recognizability, one row per key with enough citation tokens: the key's template is rebuilt without one held-out token, the token is scored against that template and against every template in the key's confusion set. n probes per key; self — median score against its own rebuilt template; top1 — percentage of probes where the own key outranks every rival; margin — median score gap to the best rival. Low top1 marks syllables whose confusion sets are genuinely hard (or whose templates are still thin) — the application uses exactly this table to warn learners which drills are unreliable.

calibration/

Raw JSON, not tabular, versioned with everything else:

• axis_norms.json — the empirical z-score distributions per scoring axis (tone, initial, sibilant, vowel, coda, …) on a native development split; the score function maps a z-score onto these percentiles. Refit in 0.4.0 after the vowel window moved.
• thresholds.json — verdict thresholds per contrast class, fit on native recordings. Refit in 0.4.0, both for the new vowel window and for a change in how the whole-syllable timbre axis is weighted; see Changelog → 0.4.0.
• style_factor.json — measured widening between read speech and citation forms, applied when transferring Common Voice widths onto citation templates.
• speaker_pitch.json — reference f0 per corpus voice; the divisor behind f1_over_f0 and f1_onset_over_f0.
• context_norms.json — how connected speech bends a syllable away from its pooled template. curves holds, for 76 combinations of the syllable's own tone and the tones on either side of it, the mean 16-point deviation of the observed contour from the template contour, measured over 28,744 corpus tokens in multi-syllable recordings. duration holds the median voiced length relative to the template by position in the phrase: first ×1.00, middle ×0.905, last ×1.16. Both are read as context-specific references, not as corrections applied to a learner's measurement.

Usage

from datasets import load_dataset

syllables = load_dataset("taiwan-corpora/twsyllables", "syllables", split="train")
templates = load_dataset("taiwan-corpora/twsyllables", "templates", split="train")
tokens    = load_dataset("taiwan-corpora/twsyllables", "tokens",    split="train")

Add revision="v0.2.0" to any of these calls to pin a release rather than track main; the tags are listed under Versions.

Taiwan tone 3 as a low tone, straight from the data:

t = {r["key"]: r for r in templates if r["style"] == "citation"}
print(t["ma3"]["tone_contour"]["center"])   # ㄇㄚˇ falls and stays low — no final rise

The ㄆ/ㄅ (p/b) aspiration contrast on the raw tokens:

import statistics
vot = lambda init: [r["vot_ms"] for r in tokens
                    if r["vot_reliable"] and r["key"].startswith(init)]
print(statistics.median(vot("pa")), statistics.median(vot("ba")))   # ㄆㄚ vs ㄅㄚ

The Parquet configs are also directly queryable — no download, no datasets dependency:

import duckdb
duckdb.sql("""
  SELECT structure.initial AS initial,
         median(vot_ms.median) AS vot_median_ms
  FROM 'hf://datasets/taiwan-corpora/twsyllables/templates.parquet'
  WHERE style = 'citation' AND vot_ms IS NOT NULL
  GROUP BY 1 ORDER BY 2
""").show()

(structure.initial is pinyin-romanized; the bopomofo spelling of each initial sits in vot_norms.zhuyin, and of each syllable in templates.zhuyin.)

import polars as pl
tokens = pl.scan_parquet("hf://datasets/taiwan-corpora/twsyllables/tokens.parquet")
tone3 = (tokens.filter(pl.col("key").str.ends_with("3"))
               .select("key", "tone_curve", "f0_ref_hz").collect())

Minimal z-scoring of one measured feature against a template:

def zscore(value, stat):
    sigma = stat.get("sigma") or max(stat["mad"], 0.8 * stat["sd"])
    return (value - stat["median"]) / sigma

tpl = t["bu4"]   # ㄅㄨˋ
z = zscore(14.9, tpl["vot_ms"])   # a measured VOT of 14.9 ms → well inside the norm

The full scoring system — contour comparison under DTW, per-axis weighting by contrast, verdict thresholds, speaker normalization from a warm-up — is what the forthcoming twspeech gem implements; the calibration files above are its exact runtime inputs. Until it ships, treat scoring reconstructions from this page as approximations.

A caution about clip-level reuse: features here were extracted by one specific implementation (window sizes, mel filterbank, YIN thresholds, LPC order). Comparing them against features from librosa/Praat without a parity check will mix measurement conventions; a published extraction specification with golden audio fixtures (from the redistributable OGDL and CC0 sources) is planned to make cross-implementation verification mechanical.

Accuracy

Two figures are reported, because they answer different questions.

Held-out syllables, corpus voices. Syllable types withheld from template fitting, spoken by the same four corpus voices the templates are built from.

Held-out speakers. The same measurement on 273 Common Voice zh-TW speakers (6,549 clips) whose recordings enter no template, no axis norm and no threshold. This is the figure that describes an unfamiliar voice.

	held-out syllables	held-out speakers
top-1 against the confusion set	0.669	0.377
tone axis, matched vs. confusable	d′ 1.35 / AUC 0.871	d′ 0.66 / AUC 0.704
initial axis, matched vs. confusable	d′ 1.20 / AUC 0.835	d′ 0.83 / AUC 0.730
final axis, matched vs. confusable	d′ 0.42 / AUC 0.663	d′ 0.19 / AUC 0.559
native tokens scored green	79.3%	53.4%
native tokens scored red or worse	1.5%	3.5%

The top-1 rows are not comparable to 0.2.0. Until 0.3.0 a rival whose template was missing for the position-specific style was silently dropped from the candidate set, which removed 23% of the rivals on average and made the task easier than it looked. The candidate set is now complete, so top-1 is measured against 4.89 rivals per syllable rather than 3.76. The axis d′ and AUC rows are unaffected by that change and improved on both columns.

The gap between the two columns is speaker generalization, and it remains the dominant error term. Four interventions were measured against it and none closed it: folding 218 of the 273 Common Voice speakers into the template centers, speaker calibration from a warm-up recording of pitch and third-formant reference, alternative rules for combining axis scores, and scaling the between-speaker term of the tolerance model between 0 and 1 (top-1 45.2% at 0, 45.6% at 1 over 12,662 held-out syllables). Each was measured against its own control on the build it was tried on.

Minimal-pair decisions — whether the whole-syllable score of the correct member of a phonemic contrast beats the incorrect one — are what a pronunciation trainer acts on. Every candidate of each family is scored, not a sample of them, and each decision is counted once from either side:

contrast	held-out syllables	held-out speakers
aspiration (p/b, t/d, k/g, q/j, ch/zh, c/z)	79.1%	77.3%
sibilant place (dental / retroflex / alveolo-palatal)	88.9%	76.3%
tone (every tone the syllable is attested with)	84.6%	69.1%
nasal coda place (-n / -ng)	64.8%	60.3%

On held-out speakers aspiration separates at every place of articulation — g/k 85.7%, b/p 83.5%, d/t 75.0%, ch/zh 72.0%, j/q 77.6% — with c/z the one exception at 51.6%, on 157 decisions. Within the sibilants the alveolo-palatal series separates from the dentals at 80.6% and from the retroflexes at 77.1%, and dental from retroflex at 75.4%, which is the de-retroflexion of Taiwanese Mandarin showing up as measurement: ㄓㄔㄕ and ㄐㄑㄒ overlap in this corpus because speakers overlap them. Nasal coda place is the weakest family on both columns — -in/-ing 71.7%, -an/-ang 58.9%, -en/-eng 51.5%; see Known limitations.

Both columns come from one command, rake pronunciation:report_card, run against this release; the tone family covers 16,756 decisions on held-out speakers and 22,641 on held-out syllables, so the tone row is the least noisy and the coda row, at 1,676, the most.

These numbers move with every pipeline change; the dataset version and MANIFEST.json checksums say exactly what you are looking at.

Validation against published measurements

The figures above say the templates separate what they are asked to separate. They say nothing about whether the underlying measurements are right. Three checks against the phonetic literature on Taiwanese Mandarin, chosen because each has a published number or a published direction to fail against.

1. Voice onset time. Two studies measured word-initial stops read by Taiwanese speakers. Our pooled medians (vot_norms) sit inside every published range, and four of the six are within 3 ms of a published mean:

initial	ours (median)	Chao & Chen 2008 (mean, range)	Chen et al. 2007 (mean, SD)
ㄅ b	11	14 (7 – 65)	13.9 (6.6)
ㄉ d	12	16 (7 – 33)	15.3 (5.7)
ㄍ g	26	27 (15 – 65)	27.4 (9.6)
ㄆ p	88	82 (35 – 147)	77.8 (23.7)
ㄊ t	89	81 (45 – 123)	75.5 (18.4)
ㄎ k	93	92 (50 – 138)	85.7 (19.4)

Our unaspirated values run 2–3 ms short of both studies, consistently across all three places — the signature of a slightly different convention for where the burst begins, not of disagreement about the contrast.

The same tables carry two patterns that Mandarin is known to break the cross-linguistic universal on, and both reproduce here. Velar VOT is far longer than labial or alveolar for the unaspirated series (ours 26 against 11 and 12), while labial and alveolar are indistinguishable — the ordering that Lisker and Abramson's "further back, longer VOT" predicts only in part. And for the aspirated series ㄊ t does not exceed ㄆ p (ours 89 against 88; published 81 against 82, and 75.5 against 77.8).

2. Tone 3 loses its rise in connected speech. Taiwanese Mandarin is described as realizing tone 3 as a low tone, with the citation-form rise absent or displaced to the syllable boundary. The position-sensitive styles let that be measured rather than asserted. Median height of the contour end above its own minimum, in semitones:

style	rise above the minimum	position of the minimum
citation	4.08 st	67% into the syllable
word	2.10 st	73%
word-initial	2.29 st	67%
word-medial	0.96 st	80%
word-final	0.90 st	80%

The full dipping tone survives only in the citation form; by word-medial and word-final position the rise is gone and the minimum has moved to the end of the syllable. Tone 4 is the control: its rise above the minimum is 0.00 st in every style, so the measure is not an artefact of the 16-point representation.

3. The neutral tone is a target, not a reduction. Huang (2018) argues the Taiwanese Mandarin neutral tone carries a static mid-low pitch target and, unlike Standard Mandarin, does not raise its pitch after tone 3. Consistent here, with the caveat that the evidence is thin: tone-5 templates have a contour range of 2.77 st against 6.09 st for tones 1–4, and of the four tone-5 cells in calibration/context_norms.json that clear the 40-token floor, the one following a tone 3 falls by 0.81 st rather than rising. One cell is a direction, not a demonstration.

References.

• Chao, Kuan-Yi & Li-mei Chen (2008). A Cross-Linguistic Study of Voice Onset Time in Stop Consonant Productions. International Journal of Computational Linguistics & Chinese Language Processing 13(2), 215–232. ACL Anthology O08-4005
• Chen, Li-mei, Kuan-Yi Chao & Jui-Feng Peng (2007). VOT productions of word-initial stops in Mandarin and English: A cross-language study. ROCLING XIX. ACL Anthology O07-2004
• Huang, Karen (2018). Phonological Identity of the Neutral-tone Syllables in Taiwan Mandarin: An Acoustic Study. Acta Linguistica Asiatica 8(2), 9–50. doi:10.4312/ala.8.2.9-50

Two further studies inform the framing rather than a specific check here: Torgerson (2005, MA thesis, Brigham Young University) reports Taiwanese Mandarin tones in a slightly lower register than Beijing; and the corpus studies of spontaneous Taiwanese Mandarin at arXiv:2503.23163 and arXiv:2606.26360 report word-specific pitch signatures and argue the neutral tone is lexical.

The word-specific finding was tested against this corpus and did not replicate, which is why templates pool all words sharing a key. A one-way decomposition of the tone contour by word over 327 keys and 17,938 Common Voice tokens — every word group attested by at least three distinct speakers — gives η² 0.079 against a permutation null of 0.059, a median excess of +0.002 per key, and −0.001 once position is held constant. The control settles it: f1_ratio and f2_ratio describe a vowel that is identical across all words sharing a key, so a word effect there is impossible by construction, and they show the same excess. The residue is small-group bias, not signal. The published effect is measured on spontaneous dialogue and against the tone contrast itself; the question here is within-tone variation in read speech, which is a much smaller quantity.

Known limitations

• Register. All token material is read speech: dictionary citation forms, dictionary words and textbook narration. No spontaneous speech, no conversation. The 全字庫 voices in particular are hyper-articulated (roughly 2× slower, narrower tone range than word-embedded speech), which is why template centers are position-sensitive and 全字庫 supports segmental, not temporal, norms.
• Speaker base of the centers. Template centers rest on 4 institutional voices; the 271-speaker diversity enters through the variance widths, not the centers. Many keys have confidence: low and a single speaker — the provenance block is there to be read. Adding speakers to the centers was measured and did not help (see Accuracy): between-speaker variance widens the template rather than sharpening it.
• Verdict thresholds are fitted on corpus voices. At the published cut 79.3% of native productions by the fitting voices land in the green band against 53.4% by voices the templates have never heard, and 1.5% against 3.5% score red or worse. Refitting the cut on unfamiliar voices was measured: it raises their green share but admits far more wrong productions, which is worse for a trainer. No cut is good on both counts, because the scores themselves separate less well on unfamiliar voices — tone AUC 0.871 against 0.704, initial 0.835 against 0.730.
• The native band covers about half the inventory, and rests on one voice. tone_contour.low and high are measured only where at least six curves survive outlier and octave-jump rejection: 804 of 1,491 word-style templates, 573 of 1,452 word-initial, 507 of 983 word-medial, 541 of 983 word-final and 778 of 1,491 citation. The remaining keys publish center and sigma alone. n_band_speakers says how many speakers stand behind each band, and for 788 of the 804 word-style bands the answer is one: the band is that speaker's spread across repetitions, widened by the between-speaker model, not a spread observed across speakers. The widening is small because the model's contour term is itself small — every token curve is normalized to its own median f0, which pins the middle of the curve near zero and leaves little between-speaker variance for the estimate to find (0.06 semitones at the middle against 0.31 at the edges). Treat the band as one native's corridor.
• The vowel window is still a fixed fraction of the rime, chosen by the rime's own shape rather than by the token's trajectory. 0.4.0 reads the nucleus later than the medial and stops before the murmur where a nasal closes the rime, which is what -ian, -ie and -ia needed — the gap between f1_vowel and f1_mid in -ia falls from 65 Hz to 1, in -ie from 54 to 16 and in -ian from 26 to 13, while monophthongs move by 2 Hz or less. A window placed on each token's own formant trajectory was tried and does not work: the maximum of the first formant is a biased order statistic that lands on the tracker's errors, and every adaptive rule measured — peak, smoothed peak, steadiest stretch, openness against flatness — scored below the fixed one on both the nucleus and the medial contrasts. What remains unfollowed is the falling diphthong: in -ai and -ao the nucleus comes first, so a window at 30–70% of the rime reads part of the glide.
• Nasal coda place is not measurable from these fields. -an/-ang and -en/-eng differ chiefly in the nucleus, not in the murmur: template F2 medians separate them by 700–1,100 Hz (ban1 2378 vs. bang1 1302 Hz; fan4 2095 vs. fang4 1172 Hz), and the labial-initial rounding of -eng toward [ʊŋ] is present as well (feng1 1043 Hz against zheng4 1538 Hz). nasal_ratio_tail carries none of it (ban1 0.72 vs. bang1 0.77). The place decision reaches 64.8% on held-out syllables and 60.3% on held-out speakers, against 79.1% and 77.3% for aspiration measured the same way, and it is the one family where reading the nucleus more accurately does not help much: -en/-eng sits at 51.5% on unfamiliar voices, barely above chance. The fields are published; the place judgment is still not one this dataset supports, and the merger is partly in the speech itself — -in/-ing and -en/-eng are reported as merging in Taiwanese Mandarin.
• Taiwan norm by design. Tone 3 is low-falling without the citation-form rise; retroflex/dental sibilant boundaries reflect Taiwanese usage. Scoring mainland Putonghua against these templates will flag exactly those differences.
• VOT is only measured where it exists — obstruent initials with a locatable release; vot_reliable: false rows carry a value measured from the window edge and should be filtered for phonetic work.
• No audio, no L2. This dataset is measurements of L1 speech. A consented L2 attempt corpus is a separate future project.

Used in

The dataset is the reference layer of the pronunciation trainer at taiwancards.com, where every learner recording is scored against these templates in real time — syllable scores, per-axis diagnostics and drill selection all derive from the numbers published here. The measurement pipeline and this dataset are developed together, which is why releases are frequent and numbers keep improving.

Related resources

• taiwan-corpora/twngrams — Taiwanese Mandarin web n-gram frequencies with host-dispersion counts, CC0.
• taiwan-corpora/twfilter-tables — the reference tables behind the twfilter variety filter.
• dsprb, dtwrb — the DSP and DTW implementations that produced every number here.

Versions

version	date	tag	DOI	revision	status
0.4.0	2026-08-25	v0.4.0	10.57967/hf/10118	5f28158	current
0.3.0	2026-08-22	v0.3.0	10.57967/hf/10102	5abec86	superseded
0.2.0	2026-08-22	v0.2.0	10.57967/hf/10094	c417a52	superseded
0.1.0	2026-08-03	v0.1.0	10.57967/hf/9812	e0b0265	superseded

Every release is a git tag in this repository and carries its own DOI. Pin the tag to reproduce a result; cite the DOI of the version the result was produced with. MANIFEST.json records the version, the row count and a SHA-256 checksum for every published file, so a working copy can be identified and verified without the Hub. date is MANIFEST.generated, the day the release was built.

revision is the commit each DOI was minted against and is what that DOI resolves to. For 0.1.0, 0.3.0 and 0.4.0 it is the tagged commit. For 0.2.0 the tag sits one commit later, on the change that recorded the DOI in this file; the two revisions differ in README.md alone and every published data file is byte-identical between them.

0.3.0 is superseded for the vowel group. Its f1_vowel and f2_vowel are read through one window fixed at points 2–3 of the eight-point track, which in a rime that opens with a medial falls on the glide: re-measured on this release's material, f1_vowel sits 65 Hz below f1_mid in -ia, 54 in -ie and 42 in -uo, where a nucleus cannot be lower than the syllable's midpoint. 0.4.0 reads the medial and the nucleus separately and closes those gaps to 2, 16 and 18 Hz. Every other field group is unaffected, and the revision and its DOI stay resolvable so that results published against 0.3.0 remain reproducible.

0.1.0 and 0.2.0 are superseded. Neither publishes a nucleus formant measurement free of nasal-murmur contamination: their only nucleus values — f1_mid, f2_mid and the ratios derived from them — are read at the midpoint of the voiced span, which in a rime closed by a nasal falls inside the murmur. Against the steady-part values published in 0.3.0, first formant is understated by 143 to 293 Hz in the open nasal rimes -an, -en, -ang, -eng and -ong, and by less as the nucleus closes. Monophthongal open rimes are unaffected, as is every field outside the vowel group. 0.1.0 is further superseded by the keying corrections listed under 0.2.0. Both revisions and their DOIs remain resolvable, so that results published against them stay reproducible. New work should use 0.3.0.

Changelog

0.4.0 — 2026-08-25

Change. The rime is measured through two windows instead of one, placed on the analysis frames rather than on the eight-point resample: the medial over frames 5–25% of the voiced span, the nucleus over 30–70%, shortened to 25–60% where structure.nasal_coda is true. Affects f1_vowel, f2_vowel, f1_onset, f2_onset and every ratio derived from them, and the templates and calibration fitted on those fields.

Motivation. 0.3.0 read the whole rime through one window at points 2–3 of the eight-point track. In a rime that opens with a medial, that window samples the glide: f1_vowel fell below f1_mid, which is impossible for a nucleus, and 0.3.0 published the discrepancy as a known limitation.

Measurement validity. Token medians of f1_vowel − f1_mid, grouped by structure.final, re-measured on this release's material under both windows. The residual is the diagnostic: a nucleus should not sit systematically below the syllable's midpoint, and the controls show what the noise floor looks like.

rime	0.3.0, Hz	0.4.0, Hz
-ia	−65	−2
-ie	−54	−16
-uo	−42	−18
-ua	−40	−4
-ian	−26	−13
-i (control, no medial)	−1	−2
-u (control, no medial)	−2	−3

F1(-ang) / F1(-a) — two low nuclei that differ little in height — is 0.91, against 0.94 in 0.3.0.

Discrimination. Two full rebuilds differing only in this code, evaluated with rake pronunciation:report_card VOICES=held_out on the 273 held-out Common Voice speakers. Every metric improves or holds except tone decisions, which lose 0.3 pp on 16,757 comparisons — on the order of the sampling error, and on an axis whose own AUC is unchanged to four decimals:

metric	n	0.3.0	0.4.0	Δ
final axis, AUC	1,936	0.5420	0.5586	+0.0166
final axis, d′	1,936	0.15	0.19	+0.04
initial axis, AUC	4,500	0.7221	0.7303	+0.0082
initial axis, d′	4,500	0.80	0.83	+0.03
tone axis, AUC	11,395	0.7038	0.7038	0.0000
aspiration, decision accuracy	2,939	76.7%	77.3%	+0.6 pp
sibilant place, decision accuracy	2,371	75.7%	76.3%	+0.6 pp
nasal coda place	1,676	59.9%	60.3%	+0.4 pp
tone, decision accuracy	16,757	69.4%	69.1%	−0.3 pp
native tokens scored green	6,371	52.8%	53.4%	+0.6 pp
native tokens scored red or worse	6,371	3.9%	3.5%	−0.4 pp
top-1 over the confusion set	6,371	0.3769	0.3769	0.0000

The gain concentrates on the final axis, which is where the misplaced window was doing its damage, and carries over to the initial axis because the medial window also feeds f2_onset_ratio, the cue that separates the sibilant series.

Rejected alternatives, both measured. (1) Placing the window on each token's own formant trajectory — first-formant maximum, smoothed maximum, steadiest stretch of the second formant, openness weighed against flatness — scored below a fixed window on both nucleus and medial contrasts, at frame and at eight-point resolution. The maximum of a noisy LPC track is a biased order statistic that selects the tracker's errors. (2) A single 30–70% window, without the nasal shortening, scored 1.2 pp higher on nasal-coda decisions and 0.8 pp higher on the native green share, but drove F1(-ang) / F1(-a) to 0.80: it bought discrimination by admitting murmur into a field named for the vowel, and was not published.

Also in this release.

• tone_contour.n_band_speakers is new, and the between-speaker widening of the band is now scaled by it — full width at one speaker, 0.6 at two, 0.3 at three, none beyond. Word-style band width rises from 0.93 to 1.04 semitones at the middle and 4.19 to 4.74 at the onset.
• calibration/thresholds.json and axis_norms.json refit. The thresholds also absorb a scoring change: the whole-syllable timbre axis carried full weight in every syllable while every other axis was scaled by how much it separates that syllable from its rivals. At a fixed 5% false-accept rate on unfamiliar voices, lowering that weight raised the share of native syllables inside the green band from 16.1% to 21.6% and the ROC area from 0.6655 to 0.6721, with top-1 unchanged.
• Token count 37,926 → 37,947, segmentation non-determinism on multi-syllable clips rather than a change of scope.

Check before re-using.

• Anyone reading f1_vowel, f2_vowel, f1_onset, f2_onset, f1_over_f0, f2_over_f1, f2_end_over_f1, f2_onset_ratio or f1_onset_over_f0 should re-pull: the values move, most in rimes with a medial and in rimes closed by a nasal. f1_mid, f2_mid and the *_ratio fields built on them are unchanged.
• formants_reliable is false for 3.8% of published tokens against 2.8% in 0.3.0. The nucleus window reads a different part of the rime, and more measurements fail the merged-formant test there. Filter on it as before.
• Tone contour bands are wider than in 0.3.0 by construction, not because the speakers vary more. Compare band widths across versions only with n_band_speakers in hand.
• The falling diphthong is the standing limitation: in -ai and -ao the nucleus comes first, so a window at 30–70% of the rime still reads part of the glide.

0.3.0 — 2026-08-22

The vowel is measured over the rime's steady part. Until now the formants were read at the midpoint of the voiced span. For a monophthong that is the vowel; for a rime closed by a nasal it is already inside the murmur, and the measurement came out wrong in a way that is visible without any statistics: median first formant 726 Hz for -ang against 1137 Hz for -a, when the two nuclei differ little in height. Read over points 2–3 of the eight-point track the same medians are 1019 and 1149 Hz. The correction scales with the openness of the nucleus — -ang +293 Hz, -eng +175, -ong +167, -an +165, -en +143, against -in +32 and -ing +36 — while monophthongal open rimes move by under 40 Hz (-a +12, -i −4, -u −5, -e −34). Diphthongs move by up to 140 Hz, but there the two points sample different parts of the glide and neither is more correct than the other. In -ian and -üan the new window falls on the medial rather than the nucleus and reads 49 and 17 Hz lower; a window fixed as a fraction of the rime cannot follow those two, and they are the known limitation of the measurement. Every figure here is a token median over tokens.parquet, grouped by structure.final. New fields f1_vowel and f2_vowel carry the measurement; f1_mid, f2_mid and the *_ratio fields built on them are unchanged and still published.

Formants are normalized against the speaker, not the syllable. f1_ratio and f2_ratio divide by a third-formant reference measured inside the same syllable, which carries vowel-dependent variation of its own. The new f1_over_f0 divides the vowel's first formant by the speaker's median pitch and f2_over_f1 takes frontness as a within-tract ratio, the two dimensions of the classical formant-ratio model. Over held-out speakers, deciding a nucleus contrast from the new pair against the old: 71.8% → 77.3%; a nasal-coda contrast: 64.4% → 72.7%. Both replicate on independent halves of the material.

The consonant-to-vowel transition is measured. f1_onset, f2_onset and the derived f2_onset_ratio and f1_onset_over_f0 read point 1 of the formant track. For the sibilants this is the documented cue that separates the alveolo-palatal series, and it lifts the sibilant contrast from 64.8% to 69.5% on held-out speakers.

The nasal coda is measured against the syllable's own middle. nasal_ratio_mid was published in 0.2.0 but unused; energy_tail_ratio and f2_end_over_f1 are new. Nasality at the end only means a coda if nasality in the middle is lower, and the energy fall into the murmur is a second independent cue. Coda-place decision on held-out speakers 48.8% → 51.6% on the axis; through the whole syllable it stays at 61%.

A tone is judged against the band native speakers actually occupy, not against their average. Averaging a contour throws away the thing a learner most needs to know: how much freedom the tone allows at each moment. Measured on the unwarped token curves, that freedom is wildly uneven — for the word-style templates the central 80% of native productions spans 0.93 semitones at the middle of the syllable and 4.19 at the onset and 5.41 at the offset. The templates now publish that band as tone_contour.low and .high, and the scorer charges nothing for a contour inside it and only the distance beyond its nearer edge outside it. Tone decisions on held-out speakers 71.9% → 73.0%, and unfamiliar native speakers are judged as natives far more often: 37.6% → 42.1% of their syllables score green and 8.1% → 5.1% score red or worse.

The band is a trimmed core, not a range: each point drops the outer 10% on either side before the between-speaker width is added, because the corpus carries real differences in recording quality and noise and the extremes are as often the microphone as the speaker. Contours whose octave jumped are discarded before the band is measured. 806 of the 1,491 word-style templates have enough tokens to earn a band; the rest keep the older comparison against the center.

Weighting the whole comparison by that spread — tightening the middle and loosening the edges in proportion — was measured and does not work: the points where speakers vary most are also the points where the tones differ most, so dividing by the native variance suppresses the signal along with the noise. Only the band form, which asks nothing inside and charges outside, keeps both.

Tone is read in context. A tone's contour in running speech is bent by the tones around it — carryover from the syllable before is assimilatory and large, anticipation from the one after is smaller. Measured over 28,744 corpus tokens, the onset of a fourth tone sits 1.39 semitones below its template after a first tone and 2.17 below it after a third; the neutral tone reproduces the textbook pattern on its own, ending 1.36 semitones above the pooled template after a third tone and 0.75 below it after a first. calibration/context_norms.json publishes those references, and the scorer compares against the one that fits the neighbours. Tone decisions on held-out speakers 71.3% → 71.9%.

Length is read by position. The final syllable of a phrase is 41% longer than a medial one in this corpus (median voiced length 310 ms against 220 ms), which is ordinary phrase-final lengthening and not something to correct a learner about. context_norms.duration carries the factors.

Tone contours are compared as shapes. The contour was compared point by point in semitones, so a learner with the right shape and a smaller excursion was penalized twice — once by the contour, once by tone_range. The contour is now standardized within the syllable, with a floor of one semitone so that level tones are not amplified into noise, and the excursion is scored only by tone_range. Tone decisions on held-out speakers 69.5% → 71.3%; the whole- syllable figure 67.0% → 72.1%.

Formant reliability is published. formants_reliable marks tokens where the tracker merged the second and third formants or the first and second — 2.8% of published tokens. It was added to the pipeline after 0.2.0 shipped.

The empty rime is labeled. syllables.apical and structure.apical distinguish the apical vowels of ㄗㄘㄙ ([ɿ]) and ㄓㄔㄕㄖ ([ʅ]) from the close front vowel of ㄐㄑㄒ, which pinyin writes identically as i. structure.initial_ipa gives ㄏ as [h] rather than [x], which is the Taiwan value.

Evaluation. Rivals whose position-specific template was missing were silently dropped from the candidate set, removing 23% of them on average. The candidate set is now complete and top-1 is measured against a harder task; see the note in Accuracy.

Token count 37,937 → 37,926; the eleven-token difference is segmentation non-determinism on multi-syllable clips, not a change of scope.

0.2.0 — 2026-08-22

Corrections. Two defects in 0.1.0 filed measurements under the wrong key. Both are fixed. Anyone who used 0.1.0 for phonetic work on the affected keys should re-pull.

• ü written as u. 92 token measurements of lǚ, lǜ and nǚ were keyed as lu3, lu4 and nu3. The templates for lv3, lv4 and nv3 therefore rested on two tokens each — the 全字庫 syllable recording alone — and lu3/lu4 mixed two rimes. Published token counts move lv3 2 → 28, lv4 2 → 45, nv3 2 → 23.
• Syllable boundaries drawn in the wrong place. 142 token measurements inside multi-syllable words were split at the wrong point and keyed accordingly — kě‑néng as ken3 + eng2, and so on. eng2 held no correct token at all and is gone.

Together 234 of 42,009 measurements, 0.56%. A key is now derived from the reading its token was measured from, with the tone taken from the corpus so that sandhi is preserved. Nine keys also carried a bopomofo spelling that contradicted the key itself (a4 as ㄍㄚˋ, qun3 as ㄑㄩˇ); their spelling is now composed from the key and their example characters, which belonged to the misspelled reading, are null.

Not a defect but a frequent misreading, now stated in the syllables field table: zhuyin carries the citation tone and key the tone spoken, so ban2 is spelled ㄅㄢˇ — its tokens are 板 bǎn raised to bán by 3-3 sandhi.

Other changes.

• Tone scoring places the syllable in the speaker's range. The contour is normalized to the syllable's own median f0, which discards absolute height; tone 1 and a connected-speech tone 3 are both near-level once height is removed. The scorer now derives f0_register from the utterance itself — the median f0 across its syllables, corrected for the tones the utterance happens to contain — and weights it at 1.5 against the contour instead of 0.5. Minimal-pair decisions on held-out speakers: 61.3% → 67.1%. The largest gains are where the axis was weakest: tone 1 vs. 2 53.9% → 68.8%, tone 3 vs. 4 54.1% → 64.2%, tone 1 vs. neutral 33.5% → 57.7% (previously below chance).
• The coda axis no longer asserts nasal place. It averaged murmur ratio, nucleus F2 and nasal antiformant, and emitted an -n-heard-as--ng verdict from that mixture, which decided the contrast at 49.7% as it stood. The axis is now the murmur ratio alone, reports only presence and strength, and is weighted 0.5 rather than 2.0, which also removes the double counting of F2 between the vowel and coda axes.
• Held-out speakers are reported. The evaluation split cut by syllable type only, so every published figure was same-speaker. Common Voice speakers are now a held-out speaker set excluded from all fitting, and Accuracy reports both columns.
• Word-final templates added. templates gains a fifth style, word_final, for 983 keys: 6,400 rows against 5,413. Introducing it also narrowed word_medial to genuinely word-medial positions, so those templates and their provenance.style blend labels moved with it.
• Syllable spellings completed. Zhuyin for a syllable×tone key outside the inventory is now composed from the key rather than left null; syllables and templates carry a spelling for all 1,491 keys. Two corrections against 0.1.0: eng1 ㄋㄥ → ㄥ, r5 ㄦ → ˙ㄦ.
• More tokens recovered. Multi-syllable recordings whose energy valleys the segmenter could not resolve are now segmented by the forced aligner instead of being dropped, which is where the 830 additional token measurements come from: 37,937 against 37,107.
• Every derived file was rebuilt from the corrected keys: templates, tokens, syllables, quality, variability, vot_norms and the calibration/ block. speaker_pitch.json lists only the four voices whose tokens are published.

0.1.0 — 2026-08-03

• First release.

Citation

@misc{taiwan-corpora-twsyllables,
  title        = {twsyllables: Taiwanese Mandarin syllable acoustics},
  author       = {Patin, Den},
  year         = {2026},
  version      = {0.4.0},
  doi          = {10.57967/hf/10118},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/taiwan-corpora/twsyllables}}
}

Author: Den Patin, ORCID 0009-0009-6496-382X — hi@dpat.in.

* Taiwanese Mandarin (臺灣華語; BCP 47 cmn-Hant-TW) denotes the variety of Mandarin in general use in Taiwan, as produced by its native and near-native speakers, together with the lexical, phonological and morphosyntactic properties that distinguish it from the codified standard of the ROC Ministry of Education (標準國語, Standard Guoyu) and from PRC Putonghua (普通話). The term follows Her (2009) and the naming of the Corpus of Contemporary Taiwanese Mandarin (臺灣華語文語料庫, National Academy for Educational Research); much of the phonetic literature calls the same variety Taiwan Mandarin. It is not 臺灣國語, the Hokkien-accented sociolect of that variety, nor Taiwanese Hokkien (臺語). Her, One-Soon 何萬順 (2009). 語言與族群認同：從台灣外省族群的母語與台灣華語談起 [Language and group identity: On Taiwan Mainlanders' mother tongues and Taiwan Mandarin]. Language and Linguistics 10(2), 375–419.