pretty_name: Consumer Speech Declipping — Clean Pool v2
task_categories:
- audio-to-audio
tags:
- audio
- speech
- declipping
- clean-reference
- 48khz
Consumer Speech Declipping — Clean Pool v2
The current release is in consumer-speech-v2/. It uses directly readable tar shards containing FLAC audio, rather than a compressed archive that needs extraction. The older dataset.tar.zst release remains available for historical work.
Contents
- 150,000 unique clean six-second training windows: 250 hours including explicit joins/padding.
- 512 fixed validation pairs and 512 fixed test pairs.
- Mono, 48 kHz, PCM24 FLAC.
- Training references receive live degradation after collation in the declipping project.
- No codec routes. Validation/test inputs and references are fixed.
- Bundle footprint: 70.35 GB.
- Bundle identity (SHA256 of canonical manifest): 9eff7c0e2e3297591e9fd508febcf86a162636f15a09afdfb8018609ed351a27
Files
- bundle.json: manifest, shard checksums, counts and build identity.
- plan.json: frozen source selections and codec-free render policy.
- items.parquet: strict item index, audio offsets and source lineage.
- audit.json: stored-data distribution and source exposure.
- shards/: directly readable tar files containing clean.flac and render records; fixed nonidentity evaluation examples also have degraded.flac.
Training records use identity storage to represent the clean reference once. This is not identity-only training. The live policy requests 20% preservation examples and uses nominal mild/moderate/severe/extreme weights of 20:40:15:5 among nonidentity draws. Actual measured damage is reported separately by the live training audit.
Sources and training sampling
The clean pool includes EARS, VCTK mic1, HiFi TTS, Espresso originals, and approved Lien/Meru microphone recordings. Original speaker splits are preserved. Lien/Meru retain all approved intervals after quiet/clipping exclusions; microphone views are not treated as additional speakers. Espresso is training-only and its speaker overlap is unverified.
Training draw shares are EARS 50%, VCTK 30%, HiFi 15%, Espresso 3%, Lien 1%, Meru 1%. These shares are independent of the stored window counts. HiFi fills the remaining unique pool capacity. The sampler excludes repeated speakers/performances within an effective batch and the two comparison architectures share a reproducible draw schedule.
Underlying source licenses and permissions apply. This combined release does not grant a new blanket license over its constituent recordings.
Download on Vast
After publication, pin the full Hugging Face commit SHA shown in the Files history. The uploader also writes that exact SHA to its local published.json receipt.
hf download anthonyleedom/test2 --repo-type dataset --revision FULL_COMMIT_SHA --include 'consumer-speech-v2/*' --local-dir /workspace/datasets
The bundle will be at /workspace/datasets/consumer-speech-v2. Download only this folder. No tar extraction is required. Use the declipping project's declip-data validation and prepare-comparison commands on that directory before training.