VoxPopuli-Platinum-en
收藏资源简介：
# VoxPopuli-Platinum-en **Filtered, word-timestamped English speech for ASR fine-tuning.** Derived from `facebook/voxpopuli` (en) with a proprietary multi-stage quality filtering pipeline - validated to improve performance on a held-out test set. > Full dataset available for licensing — contact **data@trelis.com**. ## Training Performance / Uplift **Platinum delivers improved in-domain training without the catastrophic OOD damage caused by raw VoxPopuli training.** To assess dataset quality, three training runs were conducted: one on the **unfiltered VoxPopuli labels**, one on **Whisper-large draft labels** (naive distillation), and one on **this VoxPopuli-Platinum-en dataset**. | | VP-test (in-domain) | FLEURS-en (clean read) | CV-scripted-en (clean read) | CV-spont-en (disfluent) | |---|---:|---:|---:|---:| | baseline `whisper-large-v3-turbo` (no training) | 9.01% | 4.95% | 8.65% | 18.59% | | Unfiltered VoxPopuli labels | 3.82% | 5.77% | 18.55% | 19.31% | | Whisper-large drafts as labels (naive distillation) | 7.13% | 5.48% | 9.88% | 13.27% | | **VoxPopuli-Platinum-en** | **3.03%** | 5.72% | **10.05%** | **14.58%** | Headline metrics (Platinum vs the alternatives): - **OOD protection.** Raw VoxPopuli training **catastrophically breaks** the model on clean-read OOD (CV-scripted 8.65% → **18.55%**, +9.9pp WER). Platinum preserves it (+1.4pp). That ~80% lower OOD regression is the filter's core value. - **Strong in-domain.** −66% relative WER on parliamentary English (in-domain): 9.01% → **3.03%**. - **−22% relative WER on disfluent conversational English** (CV-spont): 18.59% → 14.58%. - ~**2× better in-domain** than naive Whisper-large distillation (3.03% vs 7.13%) — proof that the curation is real signal, not just running ASR on audio. All metrics produced starting from 100h of raw VoxPopuli, trained with the same hyperparameters and for one epoch in each case, and evaluated with the "Whisper English" normalizer. ## Schema (per row) | field | type | description | |---|---|---| | `id` | string | row identifier (`vp_train_NNNNNN`) | | `audio` | `Audio` (16 kHz mono) | playable in the HF dataset viewer | | `text` | string | refined label (the training target) | | `voxpopuli_id` | string | original VoxPopuli row id (audit trail to source) | | `duration_s` | float | audio duration in seconds | | `language` | string | `"en"` | | `word_timestamps` | list of `{word, start, end, score}` | CTC-derived per-word timing + alignment confidence (negative log-prob; close to 0 = high confidence) | The 5-row sample in `data/sample.parquet` shows the schema in full. ## Sample row preview | id | duration | text (excerpt) | # words | |---|---:|---|---:| | vp_train_000003 | 2.4s | "These are not easy issues to resolve." | 7 | | vp_train_000033 | 8.0s | "In 2010, as a result of the AIFM Directive, one in four hedge funds…" | 19 | | vp_train_000022 | 9.2s | "…on the future of company law and is in phase with the Social Business Initiative…" | 23 | | vp_train_000001 | 14.2s | "…they attacked and removed the voices of resistance from our radio and TV stations…" | 38 | | vp_train_000002 | 19.8s | "In order to increase the preparedness at national and EU level…" | 38 | (Some rows start mid-phrase because VoxPopuli's source segmentation is phrase-boundary-based, not sentence-boundary-based. This is preserved intentionally — these rows are valid acoustic continuations.) ## Licensing - Audio source: `facebook/voxpopuli` (CC0 / no rights reserved) - This curated derivative: **commercial licence required** — the value is the filtering + timestamp work, not the audio itself - Buyers receive a non-exclusive use licence; redistribution is not permitted ## Test sets - `Trelis/VoxPopuli-en-test` — in-domain withheld test set (parliamentary) — **proprietary, human-labelled** - `Trelis/fleurs-en-test` — OOD clean read - `Trelis/cv-en-scripted-test-500` — OOD clean read (CommonVoice 25.0) - `Trelis/cv-en-spontaneous-test` — OOD disfluent (CV-Spont 3.0) The last three are publicly available (the latter two drawn from via Mozilla Data Collective); the in-domain test set is held proprietary to protect the headline benchmark integrity. ## Contact For commercial licensing or a 250h+ custom-curation engagement: **data@trelis.com**
