Kppwdfgu1/vocare-hasa

Tasks

• automatic-speech-recognition
• text-classification

Citation text from the card

**140 rows carry no speaker identity, no dialect and no age band.** They are 70 recordings, each
present twice, whose `audio_id` does not appear in the upstream label release, so no demographic
annotation exists for them. Their audio, transcripts, `sentence_id`, `intent_label` and `gender` are
unaffected — gender is recoverable from the recording identifier rather than from the labels.

They are not distributed evenly across the three configs, and the difference is the point:

| Config | Rows with no speaker | Why |
|---|---|---|
| `asr` (v1.1) | **0** | The speaker-disjoint protocol cannot place a recording with no speaker, so it excludes them by construction. |
| `v10-legacy` | **140** | Split by `Sentence_Id` only, so speaker attribution is not required. |
| `intent` | **36** | Split on unique `Sentence_Id`; 36 of the 70 survive deduplication. |

This is a published property of the benchmark, not a defect discovered at release time: the campaign
recorded these rows as its `no_speaker_id` bucket, and the corpus accounting closes exactly —
16,717 + 1,855 + 5,375 + 5,913 (dropped) + 140 (no speaker) = 30,000.

**`speaker_id` is the empty string on these rows, and deliberately not a sentinel such as
`"unknown"`.** A named value would read as a speaker who is *present*, and any speaker-clustered
statistic would then treat 140 recordings made by unrelated people as one person's — inflating that
"speaker's" weight and silently narrowing confidence intervals. The empty string is t

All datasets marked Permissive or share-alike | JSON