Datasets:

OpenVoiceOS
/

stt-sampler-v1

@@ -491,83 +491,31 @@ configs:
 # stt-sampler-v1
-A balanced, deterministic multi-lingual STT eval sample for
-[OVOS Plugin Arena](https://github.com/OpenVoiceOS/ovos-plugin-arena)
-leaderboards. It gives every registered STT locale immediate breadth-first
-coverage — 100 clips per language — instead of waiting for a full run over
-the much larger per-locale corpora the arena also benchmarks against
-(MInDS-14, Speech-MASSIVE, FLEURS).
-Target is 100 clips per language. Locales covered by a single source get
-all 100 from that source; locales covered by two sources (de-DE, es-ES,
-fr-FR, ko-KR, nl-NL, pl-PL, pt-PT, ru-RU) get 50 from each so no single
-source dominates a shared locale.
-Clips are selected deterministically: each source split is sorted by its
-own sample id, then strided (`index = i * floor(n / k)` for i in 0..k-1,
-k=50 or 100) so the picks spread across speakers/sessions instead of
-clustering on the corpus's first few rows (which tend to share one
-speaker). No randomness is used. Unavailable/corrupt rows are skipped by
-advancing to the next candidate in sorted order, deterministically.
-Audio is resampled to 16kHz mono float32 (MInDS-14 is native 8kHz;
-Speech-MASSIVE and FLEURS are native 16kHz).
-Total clips: **2000**
-## Sources and licenses
-| Source | HF id | License | Domain |
 |---|---|---|---|
-| MInDS-14 | [PolyAI/minds14](https://huggingface.co/datasets/PolyAI/minds14) | CC-BY-4.0 | banking (telephone intent calls) |
-| Speech-MASSIVE | [FBK-MT/Speech-MASSIVE-test](https://huggingface.co/datasets/FBK-MT/Speech-MASSIVE-test) | CC-BY-NC-SA-4.0 | assistant (spoken voice-assistant utterances) |
-| FLEURS | [google/fleurs](https://huggingface.co/datasets/google/fleurs) (ca_es, gl_es) | CC-BY-4.0 | read-speech |
-## Columns
-| column | type | description |
-|---|---|---|
-| `sample_id` | str | `<source>/<orig_id>` — source-tagged, traceable back to the origin row |
-| `audio` | `Audio` | embedded audio, resampled to 16kHz mono float32 |
-| `transcript` | str | reference transcript |
-| `lang` | str | BCP-47 locale tag |
-| `domain` | str | `banking` / `assistant` / `read-speech` |
-| `source_dataset` | str | origin HF dataset id |
-Stored as parquet with embedded audio, split by `lang` config (20
-configs) so the arena runner can resolve one HF split per language via
-`source.subset="{lang}"` in the registry entry.
-## Per-(source, language) counts
-| source_dataset | lang | domain | n |
-|---|---|---|---|
-| FBK-MT/Speech-MASSIVE-test | ar-SA | assistant | 100 |
-| FBK-MT/Speech-MASSIVE-test | de-DE | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | es-ES | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | fr-FR | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | hu-HU | assistant | 100 |
-| FBK-MT/Speech-MASSIVE-test | ko-KR | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | nl-NL | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | pl-PL | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | pt-PT | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | ru-RU | assistant | 50 |
-| FBK-MT/Speech-MASSIVE-test | tr-TR | assistant | 100 |
-| FBK-MT/Speech-MASSIVE-test | vi-VN | assistant | 100 |
-| PolyAI/minds14 | cs-CZ | banking | 100 |
-| PolyAI/minds14 | de-DE | banking | 50 |
-| PolyAI/minds14 | en-AU | banking | 100 |
-| PolyAI/minds14 | en-GB | banking | 100 |
-| PolyAI/minds14 | en-US | banking | 100 |
-| PolyAI/minds14 | es-ES | banking | 50 |
-| PolyAI/minds14 | fr-FR | banking | 50 |
-| PolyAI/minds14 | it-IT | banking | 100 |
-| PolyAI/minds14 | ko-KR | banking | 50 |
-| PolyAI/minds14 | nl-NL | banking | 50 |
-| PolyAI/minds14 | pl-PL | banking | 50 |
-| PolyAI/minds14 | pt-PT | banking | 50 |
-| PolyAI/minds14 | ru-RU | banking | 50 |
-| PolyAI/minds14 | zh-CN | banking | 100 |
-| google/fleurs | ca-ES | read-speech | 100 |
-| google/fleurs | gl-ES | read-speech | 100 |

 # stt-sampler-v1
+A small, balanced, representative multilingual ASR eval sampler for the
+[OVOS Plugin Arena](https://github.com/OpenVoiceOS/ovos-plugin-arena):
+**100 clips per language x 20 locales = 2000 clips**, 16 kHz mono float32,
+one config per locale (`load_dataset("OpenVoiceOS/stt-sampler-v1", "<lang>")`).
+Designed to seed every STT leaderboard breadth-first; full-size datasets
+accrete over time.
+## Composition
+Per locale: 100 clips, drawn 50/50 when two sources cover the locale, else
+100 from the single source. Sampling is deterministic (fixed seed) and spread
+across each source split.
+| source | license | domain | locales |
 |---|---|---|---|
+| [PolyAI/minds14](https://huggingface.co/datasets/PolyAI/minds14) | CC-BY-4.0 | banking voice queries | cs-CZ, de-DE, en-AU, en-GB, en-US, es-ES, fr-FR, it-IT, ko-KR, nl-NL, pl-PL, pt-PT, ru-RU, zh-CN |
+| [FBK-MT/Speech-MASSIVE-test](https://huggingface.co/datasets/FBK-MT/Speech-MASSIVE-test) | CC-BY-NC-SA-4.0 | voice-assistant commands | ar-SA, de-DE, es-ES, fr-FR, hu-HU, ko-KR, nl-NL, pl-PL, pt-PT, ru-RU, tr-TR, vi-VN |
+| [google/fleurs](https://huggingface.co/datasets/google/fleurs) | CC-BY-4.0 | read speech | ca-ES, gl-ES |
+Columns: `sample_id`, `audio` (16 kHz mono), `transcript`, `lang`, `domain`,
+`source_dataset`.
+## Arena integration
+Fighter predictions for this corpus are published to
+[`OpenVoiceOS/ovos-stt-bench-stt-sampler-v1`](https://huggingface.co/datasets/OpenVoiceOS/ovos-stt-bench-stt-sampler-v1)
+as `predictions/<lang>/<competitor_id>.jsonl` rows (see the arena's
+`docs/SPECIFICATION.md` §3.2).