This is the emolia-balanced-5M-subset corpus repackaged for high-quality
audio–text contrastive training. Audio is re-encoded as mono FLAC at 48 kHz
(PCM 16-bit) and stored as a WebDataset of paired .flac + .json
samples.
The JSON sidecar carries the full annotation stack:
Original metadata (id, text, duration, speaker, language, dnsmos).
A free-text emotion_caption derived from the emotion-annotation scalars.… See the full description on the dataset page:
https://huggingface.co/datasets/VoiceNet/emolia.
