kejue-countries-cities

Synthetic speech for place-name entity recognition: 6.57 h / 7,730 clips across 24 configs (19 Arabic dialects + 5 English varieties). Each clip is a phone-call utterance containing a city or country name, paired with a same-speaker continuation that names the place and the country it is in.

The continuation is the point of the dataset. A model that hears "Melbourne" should be able to place it, so every continuation contains "City, Country" verbatim:

field	example
text	"Right, so that's the branch in London, is it?"
continuation	"I need to make sure it's definitely the one in London, the United Kingdom, because I can't fly out again so soon."

Arabic configs carry Arabic renderings throughout (ملبورن، أستراليا), with the Arabic comma U+060C. City spellings are cached, so a given city is spelled identically everywhere it appears.

Schema

id, audio (16 kHz mono WAV), text, continuation, dialect, source. source encodes the generation parameters: kejue-geo/{config}/{kind}/{cc}/voice=/rate=/aug=.

How places were chosen

Ground truth is never model-generated. geonamescache supplies the city → country fact; the LLM only writes the sentence frame around an ⟨ENT⟩ placeholder and never names a country itself.

Selection is three layers, because no single signal works:

1. alternatenames (how many languages name a place) removes municipal districts — they all sit at 1 alternate name. This is what excludes Fatih/Esenler (Istanbul), Puxi (Shanghai), Bao'an (Shenzhen), Business Bay (Dubai), Shivaji Nagar (Pune).
2. A notability gate tiers each place world / national / drop. Needed because alternatenames cannot rank fame: Giza is 23 while Al Fayyum is 38, and Fes is 15.
3. A within-country gate asks which places belong among a country's best-known. Judging candidates against each other inside one country calibrates far better than judging them in isolation, which had lumped Port Said with Esna and Birmingham with Dera Ghazi Khan.

Both gates are cached to disk, so selection is deterministic and reproducible.

Resulting pool: 1,745 places / 241 countries (343 world-famous + 1,402 country-major).

Quotas per config: 15 % home country, 20 % MENA, 65 % rest of world. Every config — Arabic dialects included — draws from 71–118 different countries, at 93–99 % world tier. Measured over the shipped data: 488 distinct places across 166 countries, no place exceeding 80 of 7,730 rows.

Audio diversity

Each clip is its own TTS take (no re-use across variants). The best-known places recur under different voices, speeds and channel conditions.

• 29 voices
• 6 speeds — 1.0 (×2), 1.15, 1.3, 1.4, 1.5, pitch-preserving atempo; ~33 % unmodified
• 16 telephony conditions, weighted: clean 22 %, voip_mobile 13 %, voip_good 12 %, pstn_landline 9 %, packet loss (light/med/PLC) 12 %, AMR-NB, μ-law, bandpass, 8 kHz resample, pink noise, gain, congested/bad-signal/noisy-room VoIP

All clips are level-normalised and padded to exactly 100 ms of silence per side.

Known limitations

• ar-tn (0.16 h), ar-ma (0.20 h) and ar-lb (0.20 h) are smaller than the rest: frame generation for those dialects drops more sentences than others.
• Countries are spoken with the article where English needs it ("the United Kingdom", "the Netherlands"); city-states collapse to a single name rather than "Singapore, Singapore".
• Synthetic speech only — no human recordings.