twsyllables — Taiwan Mandarin syllable acoustics

Per-syllable acoustic reference data for Taiwan Mandarin (臺灣華語, cmn-Hant-TW): 37,107 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 Taiwan 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 Taiwan Mandarin otherwise live scattered across individual papers, usually a handful of speakers per study, and Beijing-normed references do not transfer: Taiwan 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.

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 revision (revision="..." in load_dataset) 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	5,413	acoustic reference templates: 1,491 citation + 1,491 word + 1,447 word-initial + 984 word-medial
tokens	37,107	per-recording syllable measurements: ~30 acoustic features each, with provenance
variability	22	between-speaker dispersion per feature, estimated over 271 Common Voice speakers
vot_norms	17	pooled voice-onset-time distributions per initial
quality	963	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) 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 37,107 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 18,922, moe_char 5,247, dangdai 10,014, 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 resampled to 8 points; 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. Four template styles per syllable×tone key, built from citation forms, in-word occurrences, word-initial and word-medial 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
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

templates

One row per key×style; 5,413 rows across four styles (citation, word, word_initial, word_medial). 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}	phonological parse of the syllable
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, semitones), tone_range, tone_slope, f0_register	contour is relative to the token-level reference f0
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_mid, f2_mid, f3_mid, f1_ratio, f2_ratio, f2_end_ratio, f2_delta_ratio	*_ratio fields are vocal-tract-normalized against a third-formant reference
coda	nasal_ratio_tail, nasal_ratio_mid, nasal_antiformant, centroid_ratio
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 27 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	the same feature set as templates: durations, f0_ref_hz, tone range/slope, VOT (+vot_reliable bool), frication moments, formant mids/ratios, nasal measures
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.

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.
• thresholds.json — verdict thresholds per contrast class, fit on native recordings.
• 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.

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")

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

The pipeline is validated on a held-out test split of native recordings never used for template fitting. Current figures (2026-08-03): the initial axis discriminates matched from confusable syllables at d′ 1.04 and the tone axis at d′ 1.16; 61.2% of native test tokens score green and 6.3% red under the production thresholds; the quality config publishes the per-key leave-one-out picture behind those aggregates. These numbers move — usually up — with every pipeline improvement; the dataset version and MANIFEST.json checksums say exactly what you are looking at.

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.
• 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 — Taiwan 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.

Citation

@misc{taiwan-corpora-twsyllables,
  title        = {twsyllables: Taiwan Mandarin syllable acoustics},
  author       = {Patin, Den},
  year         = {2026},
  version      = {0.1.0},
  doi          = {10.57967/hf/9812},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/taiwan-corpora/twsyllables}}
}

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