ChristophSchuhmann/levo2-vocals

LeVo 2 Vocals — 8,582 captioned a cappella singing clips

8,582 synthetic solo-singing / a cappella clips generated with LeVo 2 (gen_type: vocal), each with a detailed prose caption written by Gemini 3 Flash from the audio itself, plus structured vocal-texture flags, a music-aesthetics score and a popularity score.

The set targets a slice that is hard to find in the wild at scale: lead singing voice with little or no accompaniment, across 85 vocal genres. The distinguishing feature is not that every clip is perfectly a cappella — it is not — but that the singer-count and instrument-presence facts are labelled per clip from the audio, so you can select the subset you actually want instead of trusting the generator's intent.

These are synthetic vocals with known, documented failure modes. Read What this data is not before you train on it.

Two caveats big enough to belong up here: 35.5% of clips have audible instrumental backing despite every prompt requesting unaccompanied voice — filter on instruments_audible == False (5,537 clips) if you need a strictly a cappella subset. And genre is what was requested, not what was produced; genre_guess is what a listener actually heard.

Contents

Clips	8,582
Shards	18 WebDataset .tar
Audio	48 kHz stereo FLAC, 28.7 s mean (3.4–32.0 s)
Genres	85
Split	8,424 train / 158 val
Total audio	22.8 GB

Sources

Source	Clips	What it is
g60	6,000	60-genre x 100-clip production run, the winning recipe applied at scale
evo_phase1	1,566	Phase-1 evolutionary prompt search over 4 pilot genres (opera, choir, pop, country)
evo_transfer	600	50-genre transfer test, base and evo arms
evo_probe_acappella	186	Ablation probe: does adding a cappella to the prompt help? (it does not)
evo_probe_negation	186	Ablation probe: do negations like no choir help? (they do not)
levo_misc	44	Early small LeVo 2 runs kept for completeness

How it was generated

All audio is LeVo 2 (lglg666/SongGeneration-v2-large) run with gen_type: vocal, which is what actually produces an unaccompanied voice — asking for it in the prompt does not work (see the probes below). The generation recipe is the winner of a 10-generation evolutionary prompt search (recipe.json in this repo):

description = "<gender>, <IN-VOCABULARY genre tag>, <free style words>"
gen_type    = vocal      temperature = 0.9    top_k = 50
cfg         = 2.0        duration    = 30 s   one [verse] block
NO negations ("no choir", "no backing vocals"), NO "a cappella"

Two ablations in this dataset are the evidence for those last two rules, and both are published here as evo_probe_* clips so you can listen for yourself:

• a cappella hurts. Adding it to an opera prompt dropped the genre-match judge score from 4.17 to 1.83 out of 5 — the model reads it as a style tag and abandons the genre.
• Negation does not work. no backing vocals scored 4.33 against a plain baseline of 4.96 — statistically indistinguishable from adding backing vocals (4.29). The model keys on the noun and drops the negation. Solo-ness comes from gen_type: vocal, not from asking for it.

How it was captioned

Every clip was sent as 16 kHz mono WAV to Gemini 3 Flash through the native v1beta/models/gemini-3-flash:generateContent endpoint, with the audio inline as inline_data and a structured responseSchema.

The prompt asks the model, as a music expert, for several sentences of flowing prose covering: whether one solo singer or backing voices/choir/doubling are audible; genre and sub-genre; vocal style and technique; timbre; register and range; emotions; whether the voice is rough / gritty / fragile / breathy / pure / strained; any audible instrumentation or background noise; and the recording character.

The audio-modality assertion. The OpenAI-compatible chat endpoint for this model silently drops the audio — it returns a confident, fluent, entirely fabricated description of a track it never heard, and the only visible symptom is a suspiciously small prompt_tokens. So every single call in this dataset was checked for an AUDIO entry in usageMetadata.promptTokensDetails with a non-zero token count, and any response without one was rejected and retried rather than stored. Every caption here is backed by 6,168,003 audio tokens actually billed (median 750 per clip). If you build on this pipeline, keep that assertion — it is the one check that separates a real caption from a fabricated one.

Structured fields

Alongside the prose caption, each clip carries filterable fields:

Field	Type	True in this set
solo_voice	bool	8,019 (93.4%)
backing_voices	bool	5,307 (61.8%)
instruments_audible	bool	3,045 (35.5%)
genre_guess	string	Gemini's own genre call, independent of the prompt
emotions	list[string]

genre_guess is deliberately not conditioned on the requested genre, so genre vs genre_guess disagreement is a usable signal for "LeVo did not produce what was asked" — that is exactly how the excluded genres below were identified.

The scores, and their known disagreement

Two stored scores, both from LAION music heads running on the laion/music-whisper encoder:

• aesthetics — laion/music-aesthetics, five SongEval experts (Coherence, Musicality, Memorability, Clarity, Naturalness), each 1.0–5.0; the stored value is their mean. Per-expert values are in aesthetics_parts. Distribution here: mean 2.2779, median 2.2361, p25-p75 2.0929-2.423, range 1.4539-4.1176.
• pop_log1p_plays, pop_log1p_upvotes — laion/music-popularity-full-ft, on the model's native log1p(count) scale. Distribution: mean 0.0641, median 0.0201, p25-p75 0.012-0.0376, range -0.0023-6.1751 / mean 0.0374, median 0.019, p25-p75 0.0151-0.0255, range 0.0044-4.4123. These are not exponentiated to play counts: the model's log-MAE is ~2.1, so an exponentiated prediction is not a number anyone should quote.

Do not filter on these scores. They are stored as metadata, and they are reported here because they are informative about the scorers, not only about the audio. On a controlled 4-way comparison of music generators judged by an LLM on 30 clips each, the aesthetic head ranked HeartMuLa 4.12 with 0/30 judge passes and LeVo 2 2.41 with 22/30 passes — an almost perfectly inverted ranking:

Arm	aesthetics	LLM-judge passes (≥4/5)
HeartMuLa	4.12	0 / 30
acestep-xl-turbo	2.77	4 / 30
acestep-xl-sft	1.70	5 / 30
LeVo 2	2.41	22 / 30

The reason is domain shift: both heads were trained on finished, produced, full-band songs. Dry unaccompanied voice looks "unfinished" to them regardless of how well it is sung. They are good at spotting broken audio and bad at ranking good a cappella. Use them as an outlier detector, not as a quality filter.

The 9 genres deliberately excluded

A 50-genre transfer test judged every genre on three axes, 0–5: S (exactly one singer), G (genre matched), I (no instruments). Nine genres failed, and they failed in a very specific way — solo and no-instrument passed; only genre failed:

Key	Genre	S (solo)	G (genre)	I (no instr.)	n judged
joik	Sámi joik	3.92	1.08	4.75	12
keening	Irish keening	3.42	1.17	4.67	12
khayal	Hindustani khayal	3.75	0.83	4.75	12
throat_tuvan	Tuvan throat singing	3.25	0.92	4.17	12
patter_song	Gilbert & Sullivan patter song	4.00	1.17	4.50	12
flamenco_sae	flamenco saeta	3.64	1.09	4.27	11
ranchera	ranchera	4.25	1.33	4.58	12
forro	forró	4.25	1.42	4.75	12
appalachian_ballad	Appalachian ballad	3.17	2.42	5.00	12

Across all nine (n=107 judged clips) the means are S 3.74, G 1.27, I 4.61 — solo and no-instrument both hold up, and only genre collapses. Per-genre G ranges 0.83–2.42 out of 5. For comparison the 41 kept genres that appear in this dataset average S 3.67, G 3.42, I 4.79 over 488 judged clips: the gap is almost entirely on G (3.42 vs 1.27), not on S or I.

(appalachian_ballad is the least bad of the nine at G 2.42; it was cut with the rest because it still sits far below the kept-genre mean.)

This is a vocabulary limit, not a quality problem. LeVo 2 ships a 27-tag genre vocabulary (pop, electronic, hip hop, rock, jazz, blues, classical, rap, country, classic rock, hard rock, folk, soul, rockabilly, reggae, experimental, k-pop, experimental pop, pop punk, rock and roll, R&B, pop rock, …). Out-of-vocabulary words are not rejected — they are ignored, and the model falls back to the nearest in-vocabulary anchor. None of these nine has a near neighbour, so throat_tuvan renders as generic indie folk and joik renders as English-language pop. Spending 100 generations each on them buys nothing, so they were cut from the production run. They remain in the taxonomy as a documented capability boundary, and their transfer-test clips are still in this dataset under source = evo_transfer so the failure is inspectable.

What this data is not

• It is synthetic. Every clip is a neural generation, not a recording of a human singer. Models trained on it will learn LeVo 2's artefacts along with its singing.
• The genre label is a request, not ground truth. genre is what was asked for. genre_guess is what Gemini heard. They disagree often, and most of all on the nine genres above.
• "Solo" is not guaranteed. gen_type: vocal biases strongly toward one voice but LeVo still adds doubling, harmony stacks and choir pads on its own. That is why the solo_voice / backing_voices flags are labelled per clip — 61.8% of clips have audible backing voices despite every prompt asking for none.
• A third of the set is not actually a cappella. instruments_audible is true for 3,045 clips (35.5%), and spot-checking those captions shows real instrumental backing — acoustic guitar, bass, drum kit, synth pads — not faint artefacts or room tone. gen_type: vocal biases LeVo 2 toward unaccompanied voice but does not enforce it, and the effect is strongest in the large g60 production run. The transfer-test judge score of I 4.79/5 was measured only on the 50-genre transfer set and does not generalise to g60. For a strictly unaccompanied subset, filter ~instruments_audible — 5,537 clips remain.
• 344 clips have an empty description. These are the evo_probe_* ablation clips: their prompt genomes derived from an evolutionary elite that was later overwritten in the search state, so the exact prompt string is not recoverable. Genre, probe arm and caption are unaffected. We left the field empty rather than reconstruct a plausible-looking guess.
• Durations vary. Nominally 30 s, but the range is 3.36–32.0 s (median 30.0 s); a few generations terminated early.
• Lyrics are shared across clips. The production run reuses a small set of lyric blocks across genres by design, so the text is not diverse and should not be used as a lyrics-transcription target.
• The captions are model-written, from a single model, at temperature 0.4. They are detailed and grounded in the audio, but they are Gemini 3 Flash's opinions and inherit its biases and vocabulary.
• 5,606 clips (65.3%) were peak-normalised. LeVo renders float32 WAV that routinely peaks above 1.0, which PCM_16 FLAC cannot represent. Those clips were scaled to 0.999 peak; the gain is recorded per clip in gain_db (0.0 for everything else) and the original peak in peak_orig, so it is documented rather than silent. This is a level change only — it avoided clipping distortion rather than causing it.

Repo layout

data/vocals-{0000..0017}.tar   WebDataset shards: {id}.flac + {id}.json   <- the data
metadata.parquet                    every field for every clip, no audio
audio/{id}.mp3                      browser-streamable previews (see below)
recipe.json, genres60.json          the generation recipe and the 60-genre spec
pipeline/                           the scripts that produced all of this

audio/ is a preview tree, not the data. Mono, 24 kHz, VBR MP3, 130 KB a clip (1.1 GB total, a 17x saving against the FLAC). It exists so a browser can stream one URL per clip, which a tar member cannot do. The corpus is unaccompanied solo voice, where mono at 24 kHz costs nothing audible on a web player — but it is lossy and downmixed, so train on data/*.tar, not on audio/. metadata.parquet describes the FLAC.

Usage

WebDataset shards, {id}.flac + {id}.json per sample:

import webdataset as wds

url = "https://huggingface.co/datasets/ChristophSchuhmann/levo2-vocals/resolve/main/data/vocals-{0000..0017}.tar"
ds = (wds.WebDataset(url)
        .decode("rand")
        .to_tuple("flac", "json"))
for audio, meta in ds:
    print(meta["genre"], meta["solo_voice"], meta["caption"][:80])

Metadata alone, no audio download:

import pandas as pd
df = pd.read_parquet("hf://datasets/ChristophSchuhmann/levo2-vocals/metadata.parquet")
# clips that really are one voice, alone, with nothing else audible
clean = df[df.solo_voice & ~df.backing_voices & ~df.instruments_audible]

The validation split

split == "val" holds out 158 clips, 1 from every genre bucket, chosen with a fixed seed. It was held out before any fine-tuning and TTS-AGI/vocal-music-whisper never saw it. Keep using it as a val split and the comparison stays honest.

Fields

Field	Description
id	Clip id, unique, matches the tar member name
split	train / val
source, run, arm	Provenance — which experiment produced the clip
genre, family	Requested genre and its bucket (proven / new / probe / …)
description	The LeVo 2 style prompt used
lyrics	The lyric block fed to the model (empty for evo clips)
seed, cfg, temperature, top_k, gen_type	Generation parameters
caption	Gemini 3 Flash prose caption
genre_guess, solo_voice, backing_voices, instruments_audible, emotions	Gemini structured fields
caption_model, caption_audio_tokens	Captioner and the billed audio tokens (the modality proof)
aesthetics, aesthetics_parts	music-aesthetics, 1–5
pop_log1p_plays, pop_log1p_upvotes	music-popularity, log1p scale
judge_solo, judge_genre, judge_noinstr, judge_heard	Transfer-test LLM judge scores, 0–5 (only on evo_transfer clips)
duration, sample_rate, channels, peak_orig, gain_db	Audio properties

Genre composition

All 85 genres by clip count

Genre	Clips
pop	540
choir	504
country	474
opera	423
punk	115
indie_pop	112
sea_shanty	112
grunge	112
playground_chant	112
pop_teen	112
jpop	112
pop_boyband	112
rock_glam	112
samba	112
quiet_storm	112
lieder	112
lovers_rock	112
afrobeats	112
metal_black	112
metal_clean_baritone	112
metal_gothic_female	112
metal_power	112
opera_baritone	112
opera_mezzo	112
sprechgesang	112
oratorio	112
belt_max	112
bel_canto	112
americana_singer_songwriter	112
reggae_roots	112
blues_delta	112
bolero_latin	112
bedroom_pop	112
beatbox	112
britpop	112
bulgarian_diaphonic	112
cabaret	112
cantorial	112
choirboy_solo	112
dancehall	112
doo_wop	112
emo	112
falsetto_pure	112
field_holler	112
andean_huayno	112
northern_soul	100
opera_verismo	100
opera_countertenor	100
opera_soprano_coloratura	100
opera_tenor_dramatic	100
metalcore_scream	100
operetta_soubrette	100
motown_soul	100
surf_pop_lead	100
metal_thrash_bark	100
metal_symphonic_soprano	100
metal_doom_clean	100
metal_death_growl	100
rockabilly	100
melodie_francaise	100
brill_building_pop	100
british_invasion	100
gregorian_chant	100
rock_n_roll_50s	100
ranchera	12
throat_tuvan	12
patter_song	12
khayal	12
keening	12
joik	12
forro	12
flamenco_sae	12
appalachian_ballad	12
lantern	7
soul_jazz	3
acappella	3
rock_ballad	3
metal	3
lantern_s1234_rep	3
lantern_long	3
gospel	3
gangsta_rap	3
folk	3
lullaby	3
smoke	1

Related

The generator. The audio is LeVo 2 / SongGeneration v2. laion/moss-tts-local-transformer-4.55b-voice-acting-v2 is not the base model for anything here — that is a speech/voice-acting model and is unrelated to this music corpus.

LeVo 2 weights used (gen_type: vocal)	https://huggingface.co/lglg666/SongGeneration-v2-large
LeVo model collection	https://huggingface.co/collections/lglg666/levo
LeVo technical report (arXiv:2506.07520)	https://arxiv.org/abs/2506.07520
LeVo 2 audio samples	https://levo-demo.github.io/levo_v2_demo/

This project.

Dataset	https://huggingface.co/datasets/ChristophSchuhmann/levo2-vocals
Fine-tuned captioner	https://huggingface.co/TTS-AGI/vocal-music-whisper
Listening pages (per-genre players + scores)	https://tts-agi-moss-voice-profiles.static.hf.space/vocals/genres.html
Genre space	https://tts-agi-moss-voice-profiles.static.hf.space/genre_space.html
MOSS voice-acting manual	https://laion-ai.github.io/moss-voiceacting-manual/site/index.html

Scoring and captioning models.

laion/music-whisper (captioner base, score encoder)	https://huggingface.co/laion/music-whisper
laion/music-aesthetics	https://huggingface.co/laion/music-aesthetics
laion/music-popularity-full-ft	https://huggingface.co/laion/music-popularity-full-ft
Captioner	Gemini 3 Flash, native generateContent endpoint

Citation

@misc{levo2vocals,
  title  = {LeVo 2 Vocals: 8,582 captioned a cappella singing clips},
  author = {LAION},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ChristophSchuhmann/levo2-vocals}
}

Audio generated with LeVo 2 / SongGeneration. Captions by Gemini 3 Flash. Scores by laion/music-aesthetics and laion/music-popularity-full-ft. Released under CC-BY-4.0.