VoxEntity-EU dataset card

Composition

• synthetic_core: 180 independent scenarios (12 cells: 2 locales × 6 entity types × 15 cases), two acoustic conditions each → 360 paired audio files. Splits: 60 dev / 120 test with disjoint values and template families. Difficulty mix per cell: 5 direct, 4 orthographic/phonetic, 3 segmented/spelled, 3 containing a within-utterance self-correction.
• human_control: 50 FLEURS test rows per language (deterministic selection by sha256(fleurs:<locale>:<id>)), clean + simulated-telephone versions. Kept strictly separate from synthetic scores.

Entity types: person_name, phone, email, calendar_value, postal_address, reference_code. Every case has exactly one final target value; correction cases additionally record the superseded previous mention.

Source and generation

• Synthetic speech: Qwen3-TTS-12Hz-0.6B-CustomVoice (revision 85e237c12c027371202489a0ec509ded67b5e4b5), predefined voices serena and dylan, fp32 on a GTX 1080, sdpa attention. Masters are frozen at 24 kHz; clean is the peak-normalized 16 kHz PCM16 derivation; telephone passes the documented simulated channel (Butterworth-6 300 to 3400 Hz → 8 kHz → seeded Gaussian noise at 20 dB active-speech SNR → G.711 A-law round trip).
• The telephone condition is a simulation, not a PSTN call, and does not represent every codec.
• Human control: FLEURS de_de / fr_fr test split at the pinned revision, transformed as flagged in DATA_LICENSE.md.

Labels

Per scenario: written text, exact synthesis text, canonical value, accepted variants, mention text, difficulty tags, template family, split, seed, plus per-audio provenance (synthesis and transform records with hashes).

QA status

• Automated QA: decodability, mono, clipping fraction < 0.1%, energy, duration 1 to 45 s, abrupt-ending alerts. ASR disagreement produces alerts only, never exclusions.
• Human listening review status is tracked in IMPLEMENTATION_STATUS.md; until completed, the corpus is labeled automated-QA/provisional, not verified.

Splits

dev exists to freeze extractors and policies; test is the reporting split. Public availability of test references may contaminate future ASR models.

Extension suites (separate configs, never merged into the core tables)

Under ext/<suite>/ (gold, manifests, provenance) with their own reports:

• codeswitch: 56 DE/FR scenarios with English separators/digits, NATO spelling in German and non-DE/FR names.
• multiturn: 48 DE/FR two-turn cases (correction, confirmation, repeat); scored on the agent's persisted final state.
• voices: the 120 core test scenarios re-voiced by two more Qwen3-TTS voices.
• piper: the 120 core test scenarios synthesized by a second TTS family (Piper; French voice trained on CC BY 4.0 SIWIS data, attribution required).
• es: 90 Spanish scenarios with the core entity types and difficulty mix.

All are synthetic, with the simulated telephone channel and automated QA only.

Rights decision and takedown

Synthetic audio (Qwen3-TTS and Piper) is published by the project owner's recorded decision: Apache-2.0 at the pinned model revision with no usage or output terms found, predefined voices only, invented values; Piper voice data CC0 (thorsten) and CC BY 4.0 (SIWIS, attribution below). This is not a legal opinion. Rights or takedown requests: open an issue at https://github.com/Gjusev/voxentity-eu/issues; affected files are removed and the removal is recorded in the changelog.

Automated synthesis-failure alert (not a listening review): 7 of 180 clean core clips have both ASR systems far from both the written and the spoken text (scripts/tts_alert.py); they remain in the scored set.

Licenses and attribution

See DATA_LICENSE.md and THIRD_PARTY_NOTICES.md: annotations CC-BY-4.0; synthetic audio published by owner decision (section above); FLEURS audio CC-BY-4.0 with modifications flagged; French Piper audio carries the SIWIS attribution: Yamagishi, J.; Honnet, P.-E.; Garner, P.; Lazaridis, A. (2017). The SIWIS French Speech Synthesis Database. University of Edinburgh, CSTR. https://doi.org/10.7488/ds/1705

Intended use and limitations

• Intended: evaluation of exact spoken-entity capture (transcription, extraction, and persisted-slot perspectives) in German and French; Spanish as a separate extension suite.
• NOT intended: claims about real caller populations, causal hiring impact, production readiness, or generalization to unseen human accents. The corpus is small by design (a diagnostic, not a leaderboard); null results are published as results.
• Absolute accuracies depend on the synthesizer: the same test scenarios synthesized by a second TTS family (Piper) scored +10 pp higher with the same ASR. Treat numbers as conditional on Qwen3-TTS, not as properties of speech.
• The 360 files are 180 paired observations, not independent samples; uncertainty is reported with cluster bootstrap over scenario ids.