gradium/tts-hard-sentences-202610

TTS Hard Cases — Structured Data

A text-only evaluation set of hard cases for text-to-speech models, focused on structured data inside ordinary sentences: numbers, dates, times, codes, units, currencies and symbols. Published by Gradium.

It complements gradium/tts-eval-customer-support-202608: where that set probes long, realistic support turns, this one isolates each token type in short sentences, so a failure points at one verbalization rule.

At a glance

Rows	2,500
Unique items	500
Languages	EN, FR, ES, PT (Brazilian), DE
Categories	20
Structure	20 categories × 25 items × 5 languages
Length	2–16 words (mean 5.8 in English, 6.3–6.9 in the other languages)
Modality	Text only — no reference audio
Split	test (evaluation-only, no training split)
License	CC BY 4.0

Dataset structure

One row per (item, language). Fields:

Field	Type	Description
item_id	string	Item identifier, sd_<category>_<hash> (e.g. sd_temperature_60c45e48). Shared across the 5 languages.
category	string	Category key (e.g. slashed_dates).
category_name	string	Human-readable category name (e.g. Slashed dates).
language	string	EN, FR, ES, PT, DE.
sentence	string	The text to synthesize.
english	string	The English source sentence of the item (equal to sentence for EN).
voice	string	A suggested voice from gradium/phonon-voices.
voice_id	string	Its voice ID.

Example rows:

item_id,category,category_name,language,sentence,english,voice,voice_id
sd_temperature_60c45e48,temperature,Temperatures,EN,The temperature is -3.5°C right now.,The temperature is -3.5°C right now.,Zoey,NbpkqMVS3CJeq2j8
sd_temperature_60c45e48,temperature,Temperatures,FR,"Il fait -3,5 °C en ce moment.",The temperature is -3.5°C right now.,Noémie,FXxJ9mANRq6BCTX5

Parallel and localized

Each item_id poses the same challenge in all five languages, written from the English source in the english column with the language's own conventions: decimal commas and spaced units in French, German, Spanish and Portuguese (-3,5 °C), local date and number formats, and native phrasing rather than word-for-word translation.

Voices

Each language rotates through four voices of the Phonon voice catalogue, so a run with the suggested voices exercises several speakers per category: EN Zoey, Juniper, Everly, Harper; FR Noémie, Coralie, Tiphaine, Maëlys; ES Noa, Lucía-Sol, Candela, Celia; PT Manuela-Lu, Isadora, Débora, Rafaela; DE Nele, Jette, Annika, Jule. Any other voice works as well.

The categories

Key	Category	Items	English example
temperature	Temperatures	25	The temperature is -3.5°C right now.
ranges	Ranges	25	We are open 9-5, Monday to Friday.
alnum_codes	Alphanumeric codes	25	Your booking reference is AB-4471-Z.
slashed_dates	Slashed dates	25	Your appointment is on 14/03/2026 at 10:00.
times	Times	25	The train leaves at 07:05 from platform 3.
order_numbers	Order numbers	25	Your order number is 48213907.
files_urls	Files and URLs	25	The file is called report_final_v3.pdf.
versions	Version numbers	25	Version 3.10 fixed the login bug.
hash_numbers	Hash numbers	25	We are the #1 choice in town.
short_replies	Short replies	25	Confirmed: 14:30.
phones	Phone numbers	25	Call us on 020 7946 0018.
units	Units	25	The parcel weighs 2.4 kg.
plain_dates	Written dates	25	Your appointment is on March 14, 2026.
quantities	Quantities	25	You ordered 3 items and 2 are in stock.
mixed_fields	Mixed fields	25	Order 48213907 ships on 14/03 and arrives in 3-5 days.
currency	Currency	25	The total comes to €1,234.56.
percents	Percentages	25	You get 15% off today.
abbreviations	Abbreviations	25	Dr. Patel will see you at 10:00.
math_symbols	Math symbols	25	The margin is ±2 mm.
ordinals	Ordinals	25	This is your 3rd reminder.

Usage

from datasets import load_dataset

ds = load_dataset("gradium/tts-hard-sentences-202610", split="test")

# French dates only
dates = ds.filter(lambda r: r["language"] == "FR" and r["category"] in ("slashed_dates", "plain_dates"))
for row in dates:
    audio = my_tts_model(row["sentence"], language=row["language"], voice=row["voice"])
    save(f"{row['item_id']}_{row['language']}.wav", audio)

Or straight from the CSV:

import pandas as pd
df = pd.read_csv("data.csv")
df.groupby(["language", "category_name"]).size()

Suggested evaluation protocol

The dataset ships text only: there is no reference audio and no single "correct" verbalization — 9-5 may be read "nine to five", 14:30 "two thirty" or "fourteen thirty".

1. ASR-based proxy. Synthesize each sentence, transcribe it with a strong ASR model, and compare against the source after normalizing both sides (digits vs. words, symbols). Report the word error rate per category and per language; the per-category split is where the signal is.
2. Human scoring. For the categories where ASR normalization is unreliable (math_symbols, files_urls, alnum_codes, short_replies), have a native speaker mark the structured token correct / incorrect / ambiguous.

Limitations and scope

• Short sentences. Each item isolates one or two tokens in a short carrier sentence; long, realistic turns are covered by the customer-support set.
• No audio, no ground-truth phonemization. Scoring needs an ASR model or a human in the loop.
• Five European languages, with Brazilian Portuguese. No tonal languages, no non-Latin scripts.
• Synthetic content. Numbers, codes, names and addresses are invented; any resemblance to real records is coincidental.

License

Released under CC BY 4.0. You may use, share and adapt the dataset, including commercially, with attribution.

Citation

@misc{gradium2026ttshardsentences,
  title  = {TTS Hard Cases — Structured Data},
  author = {Gradium},
  year   = {2026},
  url    = {https://gradium.ai},
  note   = {A multilingual diagnostic set of structured-data sentences for text-to-speech}
}

About

Built and published by Gradium.