vnpost-ai/tts-vi-eval

A curated Vietnamese text-to-speech (TTS) evaluation dataset covering nine categories: ngu_am (phonology), so_dinh_dang (numeral and formatting conventions), kieu_cau (sentence type), do_dai (text length), don_vi (speech granularity), robustness (formatting stress), ngoai_lai (loanwords & foreign terms), thanh_dieu (tones) and tu_kho (hard words). Each category is published as a separate configuration (config), and every config exposes a single test split.

Overview

The dataset probes a TTS system's pronunciation accuracy across nine independent axes:

• ngu_am (phonology) — contrastive sound pairs and low-frequency phonetic structures of Northern Vietnamese: confusable onsets, voicing contrast, final-consonant pairs, vowel quality, glides (medial)/off-glides, full syllable structure, cross-syllable coarticulation, rare sounds, and aspiration.
• so_dinh_dang (numerals & formatting) — how written numbers and machine-format tokens must be rendered aloud: cardinals, ordinals, decimals, percentages, dates, clock times, monetary amounts, measurement units, scores/ratios, telephone numbers and license plates, alphanumeric identifiers, and Roman numerals.
• kieu_cau (sentence type) — sentence-final intonation by utterance purpose: declarative, interrogative (wh / yes–no), exclamatory, imperative, and conditional.
• do_dai (text length) — breath control, phrasing and prosodic stamina across four length buckets measured on text_spoken word count (short, medium, long, very_long).
• don_vi (speech granularity) — intra-utterance span: isolated syllable, word/phrase, single sentence, multi-sentence, and paragraph, probing prosody and breath grouping at each span size.
• robustness (formatting stress) — dirty input that must be read through: emoji/symbols, stray whitespace, all-caps, ambiguous word segmentation, bracketed or dashed inserts, ellipsis pauses, ultra-short and extremely long utterances, and mixed numeric fields.
• ngoai_lai (loanwords & foreign terms) — vietnamized loanwords, foreign brand, person and place names, code-switched sentences, and English graphemes read under the Vietnamese norm.
• thanh_dieu (tones) — the six tones of Northern Vietnamese: full six-tone sets, minimal and confusable tone pairs, short checked tones, pure one-tone sequences, tone co-articulation and open/nasal/stop final allophones.
• tu_kho (hard words & vocabulary features) — high-effort lexical items such as reduplicated words (tu_lay) and glottal clusters, which trip naive grapheme-to-phoneme systems.

Ground-truth read-aloud text (text_spoken) follows the Northern standard and is provided for every item, together with a pronunciation hint and, where relevant, the contrastive pairs that the sentence targets. kieu_cau, do_dai, don_vi, robustness, ngoai_lai, thanh_dieu and tu_kho instances additionally carry cross-cutting sent_*, length_* and gran_* tags in focus_tags.

Dataset Structure

Configurations and size

Configuration	Items	Sub-categories	Split
do_dai	8	4	test
don_vi	10	5	test
kieu_cau	10	5	test
ngoai_lai	12	6	test
ngu_am	18	9	test
robustness	18	9	test
so_dinh_dang	24	12	test
thanh_dieu	14	7	test
tu_kho	4	2	test
Total	118	59

Each sub-category contributes exactly two items (one easy, one hard).

Loading

from datasets import load_dataset

ds_nguam = load_dataset("vnpost-ai/tts-vi-eval", "ngu_am", split="test")
ds_so = load_dataset("vnpost-ai/tts-vi-eval", "so_dinh_dang", split="test")
ds_kieucau = load_dataset("vnpost-ai/tts-vi-eval", "kieu_cau", split="test")
ds_dodai = load_dataset("vnpost-ai/tts-vi-eval", "do_dai", split="test")
ds_donvi = load_dataset("vnpost-ai/tts-vi-eval", "don_vi", split="test")
ds_robu = load_dataset("vnpost-ai/tts-vi-eval", "robustness", split="test")
ds_ngoai = load_dataset("vnpost-ai/tts-vi-eval", "ngoai_lai", split="test")
ds_thanh = load_dataset("vnpost-ai/tts-vi-eval", "thanh_dieu", split="test")
ds_tukho = load_dataset("vnpost-ai/tts-vi-eval", "tu_kho", split="test")

Data Fields

All configs share the same 13-column schema:

Field	Type	Description
id	string	Unique instance identifier (nguam_0001, so_0001, kieucau_0001, dodai_0001, donvi_0001, robust_0001, ngoai_lai_0001, thanhdieu_0001, tukho_0001, …).
text	string	The written prompt, as delivered to the TTS system.
text_spoken	string	Reference read-aloud transcription (Northern standard).
category	string	Dataset category (ngu_am, so_dinh_dang, kieu_cau, do_dai, don_vi, robustness, ngoai_lai, thanh_dieu, tu_kho).
sub_category	string	Focus area within the category (59 values in total).
focus_tags	sequence<string>	Semantic tags describing the linguistic target, incl. sent_*, length_*, gran_* where relevant.
difficulty	string	easy or hard.
pronunciation_hint	string	Guidance on how the target must be pronounced or read.
disambiguation_hint	string (nullable)	Optional context to resolve ambiguity; unused in this release.
phonemic_ref	string (nullable)	Optional phonemic reference; unused in this release.
sensitive_pairs	sequence<sequence<string>>	Contrastive pairs the instance targets, e.g. [tr, ch].
domain	string (nullable)	Optional thematic domain; unused in this release.
source	string	Provenance; fixed to curated for this release.

Sub-category Coverage

do_dai (text length)

Sub-category	text_spoken word count	length_* tag
short	1–8 words	length_short
medium	9–20 words	length_medium
long	21–40 words	length_long
very_long	>40 words	length_very_long

don_vi (speech granularity)

Sub-category	Utterance unit	gran_* tag
am_tiet	Single isolated syllable	gran_am_tiet
tu_cum_tu	Word or short phrase (1–7 words)	gran_tu_cum_tu
cau	Single sentence (8–20 words)	gran_cau
nhieu_cau	Two to three sentences (21–40 words)	gran_nhieu_cau
doan_van	Paragraph (≥4 sentences, >40 words)	gran_doan_van

kieu_cau (sentence type)

Sub-category	Sentence type	sent_* tag
tran_thuat	Declarative	sent_tran_thuat
cau_hoi	Interrogative (wh / yes–no)	sent_nghi_van
cam_than	Exclamatory	sent_cam_than
menh_lenh_cau_khien	Imperative	sent_cau_khien
dieu_kien_nhan	Conditional	sent_dieu_kien

ngoai_lai (loanwords & foreign terms)

Foreign graphemes are kept verbatim in text_spoken; the intended reading follows the habitual Vietnamese pronunciation norms.

Sub-category	Foreign-term condition exercised
vietnamized_loanword	Loanwords fully vietnamized (tắc xi, sô cô la).
brand_product	Foreign brand names (Mercedes, iPhone, Samsung).
foreign_person_name	Foreign personal names (Bill Gates, Oliver Twist).
foreign_place_name	Foreign place names (Paris, London Heathrow).
mixed_sentence	Code-switched sentence (hi everyone, thank you).
english_grapheme_vi_norm	English graphemes read under the Vietnamese norm (card, laptop).

ngu_am (phonology)

Sub-category	Linguistic target
initial_confusable_pair	Contrastive onsets such as tr/ch, d/gi.
vo_thanh_huu_thanh	Voiceless vs. voiced plosive contrast (t/đ, c/g).
final_consonant_pair	Final-consonant contrast (c/t, n/ng).
rime_vowel_quality	Vowel quality and quantity (â/ơ, ai/ay).
glide	Medial and off-glides (qua, loà xoà).
syllable_structure_matrix	Full onset + rime + tone integrity.
cross_syllable_coarticulation	Coarticulation across syllable boundaries.
rare_sound	Low-frequency phonetic structures (ngh before i/ê, k before i/e/y).
aspirate_voiced	Aspirated consonants (th, ph, kh).

robustness (formatting stress)

text keeps the dirty formatting the TTS must read through (emoji, digits, stray whitespace, caps, punctuation). text_spoken is the clean normalized reference read-aloud and is always free of digits and emoji.

Sub-category	Stressing condition exercised
emoji_symbol	Emoji / pictographs that must be ignored in speech.
extra_space	Stray duplicated whitespace between words.
caps_uppercase	All-caps input read with normal prosody.
wrong_segment	Syllable sequences prone to word-segmentation errors.
quote_parenthesis_dash	Bracketed or dashed inserts kept fluent.
ellipsis_pause	… used to mark hesitation / pause.
short_edge	Ultra-short utterances (1–3 words).
long_edge	A single extremely long sentence (>40 words).
mixed_field	Digits, currency, codes and punctuation in one sentence.

so_dinh_dang (numerals & formatting)

Sub-category	Reading convention exercised
cardinal	Whole-number reading, including zero handling (1005).
ordinal	Ordinal reading (thứ 150 → thứ một trăm năm mươi).
decimal	Decimal point as phẩy, digit-by-digit fractional part.
percent	Percentage as phần trăm.
date	Day/month/year, including mồng một and full year.
time	Clock as hour + giờ + minutes (phút).
money	Grouped monetary amounts with đồng.
unit	Measurement units, incl. decimal amount + unit.
ratio_score	Score read as two numerals.
phone_id_plate	Digit-by-digit reading incl. letters (F8 → ép tám).
alphanumeric_mixed	Mixed alphanumeric codes (THCS0001, 12A3).
roman_numeral	Roman numerals as cardinals.

thanh_dieu (tones)

Sub-category	Prosodic target	thanh_* tag
six_tone_set	Full six-tone set on one shape (ma mà mả mã má mạ).	thanh_<ngang/huyen/hoi/nga/sac/nang>
minimal_pair_tone	Minimal tone pairs in context (bố/bộ, củ/cũ).	tone_pair_<t1>_<t2>
confusable_tone_pair	Easily confused tones in proximity (sửa/sữa).	confusable
entering_tone	Short checked tones before p/t/c/ch (mất, học).	thanh_sac_short/thanh_nang_short
tone_quality	Pure one-tone sequences for steady pitch measurement.	thanh_<tone>
tone_sandhi	Tone co-articulation in running speech (cũng ngủ).	thanh_nga/thanh_hoi
tone_final_allophone	Open / nasal / stop allophones under the same tone (má/mán/mát).	thanh_sac

tu_kho (hard words & vocabulary features)

Pilot release covers two sub-categories; more to follow in later versions.

Sub-category	Lexical feature exercised
tu_lay	Reduplicated words (lấp lánh, bâng khuâng), onset and rime.
glottal_cluster	Glottal clusters (o ạch, ạ ực) with abrupt articulation.

Intended Use

This is an evaluation-only resource; it is not intended for training. It is suited to:

• benchmarking Vietnamese TTS pronunciation fidelity against a Northern-standard reference;
• stress-testing number- and format-reading rules (word form, cardinals, dates, codes);
• probing sentence-final intonation and length-dependent prosody;
• benchmarking across utterance granularity from isolated syllables to full paragraphs;
• validating robustness to dirty formatting (caps, whitespace, emoji, punctuation);
• checking loanword and foreign-name handling under Vietnamese reading norms;
• measuring tone accuracy: six-tone sets, confusable pairs, checked tones and tone sandhi;
• probing hard lexicon items (reduplications, glottal clusters) for grapheme–phoneme fidelity;
• regression testing of spoken-text normalization pipelines.

Provenance & Maintenance

The instances are hand-curated from data/test_sets/nguam_sodinhdang_kieucau_dodai_donvi/eval_vi_tts_nguam_sodinhdang_kieucau_dodai_donvi.json and published through the build tooling in the source repository (src/eval_ds/focused_combined.py, src/eval_ds/hf_export.py). The dataset is maintained by the vnpost-ai team.

Citation

@misc{vnpost-ai-tts-vi-eval,
  title  = {Vietnamese TTS Evaluation Dataset: Phonology \& Numerals}, 
  author = {vnpost-ai},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/vnpost-ai/tts-vi-eval}}
}