Datasetpublic

Trelis/VoxPopuli-Platinum-en-full

VoxPopuli-Platinum-en-full Complete 142,066-row / ~404-hour English VoxPopuli Platinum dataset. Reach out to data@trelis.com to purchase access or discuss a larger custom-curation engagement. Training Results These results show why the Platinum labels matter. We compare the base model, fine-tuning on raw VoxPopuli transcripts, and fine-tuning on this internally filtered Platinum dataset. Evaluation uses the same english-spoken corpus WER setup across four… See the full description on the dataset page: https://huggingface.co/datasets/Trelis/VoxPopuli-Platinum-en-full.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes114downloads

Dataset Card

VoxPopuli-Platinum-en-full

Complete 142,066-row / ~404-hour English VoxPopuli Platinum dataset.

Reach out to data@trelis.com to purchase access or discuss a larger custom-curation engagement.

Training Results

These results show why the Platinum labels matter. We compare the base model, fine-tuning on raw VoxPopuli transcripts, and fine-tuning on this internally filtered Platinum dataset. Evaluation uses the same english-spoken corpus WER setup across four held-out ASR test sets.

[image]

Qwen3-ASR-0.6B WER

Training data	VP-test	FLEURS	CV-scripted	CV-spontaneous
Base model, no fine-tune	2.39%	4.40%	7.31%	9.16%
Raw VoxPopuli transcripts	4.49%	4.70%	19.45%	16.77%
VoxPopuli Platinum	2.17%	4.02%	7.73%	9.90%

Whisper Large v3 Turbo WER

Training data	VP-test	FLEURS	CV-scripted	CV-spontaneous
Base model, no fine-tune	8.26%	4.41%	8.38%	10.93%
Raw VoxPopuli transcripts	3.77%	7.08%	15.75%	24.27%
VoxPopuli Platinum	3.12%	5.90%	11.41%	11.36%

Raw VoxPopuli labels are useful in-domain but can damage out-of-domain behavior. The Platinum set is filtered with Trelis internal heuristics to keep the useful VoxPopuli signal while rejecting or repairing label/audio mismatches. The training results above are the evidence for that quality filter; the filtering recipe itself is proprietary.

—10 parquet shards: data/train-00000.parquet through data/train-00009.parquet
—142,066 rows / about 404 hours
—16 kHz English speech from VoxPopuli
—Cleaned training text
—Word-level timestamps
—Speaker IDs for every row
—Original VoxPopuli raw_text and normalized_text for auditability

Recommended Use

This dataset is suitable for ASR fine-tuning, timestamp-aware ASR experiments, captioning/alignment workflows, speaker-aware filtering, and TTS experiments. For TTS, treat it as parliamentary speech with cleaned text and speaker IDs, not as studio-recorded voice-cloning data.

Speaker IDs

source	rows	meaning
`voxpopuli`	91,482	Original VoxPopuli speaker ID was present and preserved
`inferred_matched`	16,104	Missing-source row was matched to an existing speaker identity
`inferred_cluster`	34,480	Missing-source row was assigned to an inferred speaker cluster

The speaker_id_source column tells you whether the speaker ID came directly from VoxPopuli or was inferred by Trelis.

Schema

field	description
`id`	Trelis row ID
`audio`	16 kHz audio
`text`	cleaned ASR/TTS training text
`text_source`	internal label-source marker
`voxpopuli_id`	original VoxPopuli `audio_id`
`duration_s`	clip duration
`language`	`en`
`gender`, `accent`, `is_gold_transcript`	source VoxPopuli metadata
`voxpopuli_raw_text`	original VoxPopuli raw transcript text
`voxpopuli_normalized_text`	original VoxPopuli normalized transcript text
`word_timestamps`	list of `{word, start, end, score}`
`speaker_id`	preserved or inferred speaker ID
`speaker_id_source`	`voxpopuli`, `inferred_matched`, or `inferred_cluster`

License

Commercial Trelis license required. Redistribution is not permitted.