laion/6k-diverse-reference-voices

6k Diverse Reference Voices

6,064 permissively licensed reference voices for casting expressive voice-acting generations.

All voices in this collection are permissively usable: they were either synthetically created or extracted from the CC-BY part of Emilia. Licensed under CC-BY-4.0.

Source / attribution: derived from TTS-AGI/moss-reference-voices-consolidated (CC-BY-4.0), re-published under LAION with clarified metadata documentation. If you use this dataset, please attribute both LAION and the original TTS-AGI source.

Every voice was auto-annotated by Gemini (name, tagline, language, accent, age/gender read, register, timbre, distinctive features, emotional range, casting suggestions for 4 genres, free-text tags, search text) and scored on 99 measured dimensions: 57 VoiceNet voice-quality axes, 40 Empathic-Insight emotion axes, plus genuineness and burst-blend. Each voice ships with three audio variants and per-variant DNSMOS.

Live search UI: https://projects.laion.ai/moss-reference-voice-search/ — code + reproducer under search_tool/ in this repo.

Composition

Counts by source (= per-voice source field, also the cid prefix). Verified from metadata.parquet:

source	count	what it is
emolia	3,000	Emilia-derived voices, kept with traceable id (emolia_c*).
char	1,336	Synthetic character voices (k<n>_age<n>_bg<n>).
refvoice	956	Synthetic reinterpreted reference voices, opaque id.
mediathek	472	Synthetic German reinterpretations, opaque id.
anime	300	Synthetic anime-derived reinterpretations, English delivery, opaque id.
total	6,064

Language (Gemini-identified, 4,727/6,064 non-null): English 4,250 · German 473 · null 1,337 · rest 4. Gender read: Male 4,262 · Female 1,675 · Androgynous 109 · Non-human 8 · edge cases 10.

The three audio variants — and why

Each voice has three parallel renders of the same demo clip:

variant	tar suffix	what it is
orig	<cid>.orig.mp3	Original demo clip.
sidon	<cid>.sidon.mp3	SIDON-denoised + loudness-normalized orig. Can introduce metallic ringing / over-smoothing on some clips.
cbx	<cid>.cbx.mp3	Chatterbox self-conversion of the sidon clip (artifact-cleanup pass, occasionally softens texture).

Dataset-level mean DNSMOS-OVRL is tied (orig 3.343, sidon 3.346, cbx 3.344 — see annotations/dnsmos_stats.json), but per-voice ~62% win with a processed variant (orig wins 2,317 = 38.2%, sidon 1,940 = 32.0%, cbx 1,807 = 29.8%). Use the precomputed best_version field (argmax DNSMOS per voice) instead of defaulting to one variant dataset-wide.

File layout

data/voices-0000.tar … voices-0011.tar   # WebDataset shards, ~505-506 voices each, ~2 GB total
metadata.parquet                          # flat index, one row per voice (6064 x 30, see below)
annotations/
  dims.npy            # (6064, 99) float32 — 99-dim scores on ORIG audio
  dims_enh.npy        # (6064, 99) float32 — 99-dim scores on SIDON audio
  dim_catalog.json    # 99-dim schema: [{i, code, name, group, desc}]
  dnsmos.json         # {cid: {orig, sidon, cbx}} DNSMOS-OVRL per variant
  dnsmos_stats.json   # means + per-variant win counts / win_pct
search_tool/          # FastAPI search server + pipeline + demo page
README.md / LICENSE   # this file / CC-BY-4.0

WebDataset shards (data/*.tar)

Members for one voice are contiguous:

<cid>.orig.mp3     # original demo clip
<cid>.sidon.mp3    # SIDON-denoised + loudness-normalized
<cid>.cbx.mp3      # Chatterbox self-conversion of sidon
<cid>.json         # full per-voice record (see below)

The <cid>.json record = full voice entry (name, tagline, gender, age, language, accent, register, timbre_profile, distinctive_features, emotional_range, casting {classic_fantasy, sci_fi, mystery_horror, contemporary}, tags, search_text, legacy scores, source) plus:

{
  "dnsmos": {"orig": 3.44, "sidon": 3.40, "cbx": 3.29},
  "best_version": "orig",
  "dims_raw": [99 floats, order = annotations/dim_catalog.json, scored on orig],
  "dims_enh": [99 floats, same order, scored on sidon]
}

import webdataset as wds
ds = wds.WebDataset("hf://datasets/LAION/6k-diverse-reference-voices/data/voices-{0000..0011}.tar").decode()
for sample in ds:
    cid = sample["__key__"]
    rec = sample["json"]  # dict with dnsmos, best_version, dims_raw, dims_enh
    print(cid, rec["name"], rec["best_version"])

metadata.parquet — flat index (1 row / voice)

Columns (30 total):

• Identity / casting: cid, name, gender, age, language, accent, tagline, tags (list<string>), source, shard
• Quality per variant: dnsmos_orig, dnsmos_sidon, dnsmos_cbx (float, DNSMOS-OVRL 0–5, higher = better), best_version (orig|sidon|cbx = argmax DNSMOS)
• Key 99-dim values with both scorings (dim_* = on orig, dim_*_enh = on sidon): dim_GEND (+_enh) perceived gender (higher = more masculine), dim_AGEV (+_enh) perceived age, dim_GENU (+_enh) genuineness (sounds like real human recording), dim_BLEND (+_enh) vocal-burst blend quality, dim_BKGN (+_enh) background-noise level, dim_VALN (+_enh) / dim_AROU (+_enh) emotional valence / arousal, dim_WARM (+_enh) vocal warmth.

import pandas as pd
df = pd.read_parquet("metadata.parquet")
loud_masculine = df[(df.dim_GEND > 4) & (df.dnsmos_orig > 3.4)]

Row order of metadata.parquet == row order of annotations/dims.npy / dims_enh.npy.

99-dim scores (annotations/)

• dim_catalog.json: list of 99 {i, code, name, group, desc}, i = column index into the .npy files. Groups: emonet 40 (indices 0–39, higher = stronger emotion), voicenet 57 (indices 40–96, timbre/prosody/register/style), quality 2 (index 97 GENU genuineness, 98 BLEND vocal-burst blend).
• dims.npy: float32 (6064, 99), scored on orig audio. dims_enh.npy: same shape, scored on sidon audio. No NaNs. Observed range approx −3.9 … 12.2 (raw regressor outputs, not clipped to 0–6).
• Per-tar dims_raw == corresponding dims.npy row; dims_enh == dims_enh.npy row.
• dnsmos.json: {cid: {"orig": float, "sidon": float, "cbx": float}} for all 6,064 voices.

License

CC-BY-4.0. You may share and adapt with attribution to LAION and the original TTS-AGI/moss-reference-voices-consolidated source.

Search

See search_tool/ for the FastAPI server (BM25 / sentence-embedding / VoiceCLAP text→audio similarity, optional AND-filters over any of the 99 dims), the pipeline scripts (SIDON, Chatterbox self-conversion, DNSMOS, 99-dim scoring, VoiceCLAP, assembly), and the live-demo reproducer.