Datasetpublic

leeoxiang/emilia-expressive-zh

Emilia Expressive — Phase-1 Filtered Subset Auto-generated by emilia_pipeline.scoring.phase1_hf. This is the Phase-1 filtered view: every clip that survived the S0+S1 acoustic funnel, physically partitioned into quality tiers so you can download exactly the strictness level you want -- before Phase-2 emotion labeling. Derived from amphion/Emilia-Dataset (CC-BY-NC-4.0); the same license and usage restrictions apply. Pipeline version: voxsift-emilia-v1.3-full (schema 1.3) Clips:… See the full description on the dataset page: https://huggingface.co/datasets/leeoxiang/emilia-expressive-zh.

sourceHugging Facecc-by-nc-4.0updated 3mo agoView on Hugging Face

0likes595downloads

Dataset Card

Emilia Expressive — Phase-1 Filtered Subset

Auto-generated by emilia_pipeline.scoring.phase1_hf. This is the Phase-1 filtered view: every clip that survived the S0+S1 acoustic funnel, physically partitioned into quality tiers so you can download exactly the strictness level you want -- before Phase-2 emotion labeling.

Derived from amphion/Emilia-Dataset (CC-BY-NC-4.0); the same license and usage restrictions apply.

—Pipeline version: voxsift-emilia-v1.3-full (schema 1.3)
—Clips: 4,487,838 (4,487,838 with audio)
—Format: WebDataset tar shards under data/{tier}/; each clip is {clip_id}.mp3 + {clip_id}.json (loadable with datasets.load_dataset("webdataset", ...) or the webdataset library). Audio keeps its original sample rate (Emilia-ZH mixes 24/32/44.1 kHz -- see the sample_rate metadata column); it is never resampled by the pipeline.
—Metadata: metadata/phase1_metrics.parquet -- one flat row per clip (full S0-S3 metrics + tier + sample_rate, keyed by clip_id). Phase-2 emotion/prosody labels are published incrementally as metadata/s4_labels.parquet with the same key; audio tars are never rewritten.

Tiers: pick your filtering level

Tier	Selection rule	Clips
`prime`	S3 speaker-purity pass (single, intrudedtrimmed, degradedpass) AND `prosody_dsp_score` in the global top 40% of S3-pass clips — the curated expressive core	1,300,693
`extended`	S3 pass, below the top-40% prosody cut — clean but prosodically flatter	1,951,038
`s3rejected`	S1-pass but rejected by S3 sliding-window purity (possible speaker intrusion / degradation); shipped verbatim, no trim, use at your own risk	1,236,107

Rules of thumb:

—Just want the best expressive TTS data -> download data/prime/ only.
—Want more hours, still clean -> data/prime/ + data/extended/, then optionally re-cut by prosody_dsp_score yourself.
—Custom funnel -> take all tiers and filter on metadata/phase1_metrics.parquet: every S0-S3 metric is a column, so any stricter (or looser) gate is a parquet query, no repacking needed.

Filtering funnel (applied upstream of the tiers)

Stage	What it does
S0	Metadata prefilter: 3.0-30.0s, lang=`zh`, original DNSMOS ≥ 3.2, text ≥ 4 chars
S1	Acoustic gate: aespq ≥ 7.0, aespc ≤ 2.5, aes_ce ≥ 5.0, SNR ≥ 20.0 dB, bandwidth ≥ 4000.0 Hz
S2	Prosody richness score (`prosody_dsp_score`, z-scored over all S1 survivors); the top 40% cut among S3-pass clips defines `prime`
S3	Sliding-window speaker purity; pass verdicts (single, intrudedtrimmed, degradedpass) split `prime`/`extended`, the rest -> `s3rejected`

Clips are ordered by labeling priority (prosody_dsp_score * norm_aesthetics_pq); the priority_rank field preserves that order. intruded_trimmed clips ship head/tail-trimmed audio matching their advertised duration, re-encoded as high-quality VBR MP3 (all other audio is the verbatim source bytes; s3rejected audio is always verbatim).

Per-clip JSON schema

Identity (clip_id, source_shard, text, speaker, language, duration_s), tier, purity (verdict + trim bounds + gender), and labeling priority are always present. When packaged with include_metrics, a metrics block carries the full S0-S3 acoustics + prosody numbers.