Datasetpublic

BrunoHays/minds14-telephony-pt

MINDS-14 telephony — Portuguese (pt-PT) Processed banking IVR telephony speech derived from PolyAI/minds14 (pt-PT / related config). Credits & license Original corpus: PolyAI MInDS-14Gerz et al., 2021 — arXiv:2104.08524. License: CC BY-4.0 (commercial use allowed with attribution). Snapshot Language: pt (locale pt-PT) Samples: 604 clips (~2.685 h) after GT remaster Audio: OGG/Opus mono, native 8 kHz telephony Split: all test (eval-only remaster;… See the full description on the dataset page: https://huggingface.co/datasets/BrunoHays/minds14-telephony-pt.

sourceHugging Facecc-by-4.0updated 1d agoView on Hugging Face

0likes29downloads

Dataset Card

MINDS-14 telephony — Portuguese (`pt-PT`)

Processed banking IVR telephony speech derived from PolyAI/minds14 (pt-PT / related config).

Credits & license

Original corpus: PolyAI MInDS-14 Gerz et al., 2021 — arXiv:2104.08524.

License: CC BY-4.0 (commercial use allowed with attribution).

Snapshot

—Language: pt (locale pt-PT)
—Samples: 604 clips (~2.685 h) after GT remaster
—Audio: OGG/Opus mono, native 8 kHz telephony
—Split: all test (eval-only remaster; upstream was train-only)
—Ordering: longest → shortest
—License: CC-BY-4.0

Processing applied

1.Remaster PolyAI audio to OGG/Opus mono at native 8 kHz telephony.
2.Schema columns: original_transcription, processed_transcription, transcription.
3.Upstream text is Google ASR 1-best, not human GT (see paper / Hub card annotations_creators: machine-generated). Kept immutable in original_transcription.
4.Remaster reference: processed_transcription / transcription = Gemini 3.5 Transcribe on every kept clip.
5.Drops: noise-only legacy GT stubs, empty Gemini hypotheses, and wrong spoken language (e.g. English audio in a non-English split). Dropped here: 0 clips / 0.000 h.

Dataset structure

field	description
`audio`	OGG/Opus, 8 kHz mono telephony
`audio_id`	stable sample id
`license`	`CC-BY-4.0`
`original_transcription`	upstream Google ASR 1-best (immutable)
`processed_transcription`	Gemini remaster transcript
`transcription`	best available (= processed)
`dur` / `sample_rate` / `lang` / `split`	metadata
`intent_class` / `rank` / `src`	optional extras when present

Notes

Short single-turn banking intents — useful as an 8 kHz telephony acoustic canary, not a long-form conversational bench. Contamination risk is high (widely used in training corpora).