TTS-AGI/levo2-vocals-gender-rebalance

LeVo 2 vocals — gender-rebalancing expansion

4,800 generated solo-vocal clips (39.1 h) across 60 genres, made to fix a measured male skew in TTS-AGI/levo2-vocals-captioned.

This is not a standalone corpus. It is the second half of one — the expansion that was generated after the first half was found to be 4.8 : 1 male, and it exists so that the combined pool can be downsampled to a genuinely 50/50 training set without duplicating a single clip.

The problem this set fixes

The original 8,582-clip corpus was generated from prompts that begin with a gender word. Captioning and judging it revealed the skew, and a controlled probe found the cause is prompt word order in the generator, not the captioner. 36 clips per arm, same genres, same seeds, only the position of the gender word changed:

prompt variant	female share
base — gender word first	11.1 %
female_singer	6.7 %
gender_genre	13.9 %
gender_only	13.9 %
woman_singing	18.2 %
gender_last — gender word last	47.2 %

Confirmed acoustically, not only by the judge: median F0 324 Hz for the clips the judge called female against 276 Hz for male.

Every clip here was generated with the winning gender_last variant (temperature 0.9, cfg 2.0, top-k per scaleup_jobs.json), one prompt template applied across all 60 genres, 80 clips per genre, seeds recorded per clip.

What it actually achieved — read this before trusting the fix

corpus	n	female	male	female share
original (levo2-vocals-captioned)	8,582	1,483	7,090	17.3 %
this expansion	4,800	1,740	3,057	36.3 %
combined pool	13,382	3,223	10,147	24.1 %

36.3 %, not the 47.2 % the probe promised. The probe arm was n=36; a second probe of the same variant at larger n measured 36.1 %, which is what the 4,800-clip run reproduced almost exactly. The 47.2 % figure is small-sample noise and the honest expected yield of this recipe is ~36 %. A follow-up probe found cfg 4.5 reached 44.4 % (n=36) — that setting was measured but not used for this run, and remains the most obvious untried improvement.

So the prompt-order fix roughly doubles female yield (17.3 % → 36.3 %) and does not solve the problem outright. It moves the combined pool far enough that a 50/50 subset of 6,378 clips can be drawn without oversampling; that subset is what the retrained captioner was trained on, and it is marked in this dataset.

Layout

WebDataset tar shards plus a single parquet. No loose files.

data/femvocals-{0000..0007}.tar   600 samples each: <id>.mp3 + <id>.json
metadata.parquet                  one row per clip, same fields as the json
stats.json

Audio is mono 24 kHz mp3 @ 96 kbps — this set exists to be listened to and re-captioned, not to be a mastering source. The 48 kHz stereo float32 originals stay on the cluster and source_wav records their path.

Fields

field	meaning
id, genre, duration_s, shard	identity
caption	Gemini 3 Flash description of the audio, the fine-tuning target
solo_voice, backing_voices, instruments_audible	structural flags from the same caption pass
judge_gender	majority verdict of the Gemini gender judge
judge_votes, judge_agree, judge_unanimous	how many reps voted and how they split
judge_voice_type	soprano / tenor / … as called by the judge
prompt_variant	always gender_last
prompt_description, lyrics, seed, temperature, cfg, top_k	full generation recipe
in_balanced_train	true for the 2,706 clips that entered the 50/50 training set
source_wav	path of the 48 kHz original on the cluster

The balanced training set

in_balanced_train marks this dataset's contribution to a 6,378-clip, exactly 50/50 training set (3,189 female / 3,189 male, 85 genres). It was built by:

1. no oversampling — every clip appears at most once; duplicating the scarce female clips to hit a round number produces a fake balance and a model that has memorised 500 voices;
2. stratified downsampling of the male side, genre by genre, with a floor of 12 per genre, so all 85 genres survive and the balance is not concentrated in a few buckets;
3. male clips chosen by deterministic id hash, so the selection is reproducible and not correlated with caption length or quality;
4. clips the judge could not gender are dropped rather than assigned a side.

Two instruments, and they disagree

Gender here has two independent labels and it would be dishonest to quote only one. The audio judge (Gemini listening to the clip) and the caption-derived label (token majority over the Gemini caption text) agree per-clip 84.8 % of the time, yet differ roughly 3× on the corpus-wide female base rate. This dataset publishes the judge label because it is the one derived from the audio. Treat both as instruments, not ground truth.

Limitations

• Synthetic. Every clip is a LeVo 2 gen_type: vocal generation. Expect a transfer gap to real human singing.
• Still not balanced on its own — 36.3 % female. Balance is achieved only by combining with the original corpus and downsampling, as above.
• 60 genres here, 85 in the combined set. The expansion covers the genres present in the generator's genre table; the balanced set's remaining genres are carried by the original corpus and by the per-genre floor.
• Gender is judged, not known. There is no speaker identity behind these voices; the label is one model's perception of another model's output.
• The generator's known style failures carry over: for several genres LeVo 2 provably cannot produce the requested style, and the caption describes what was generated, not what was asked.

Related

Original corpus	https://huggingface.co/datasets/TTS-AGI/levo2-vocals-captioned
Captioner retrained on the balanced set	https://huggingface.co/TTS-AGI/vocal-music-whisper
Base captioner	https://huggingface.co/laion/music-whisper
Generator (LeVo 2 / SongGeneration v2)	https://huggingface.co/lglg666/SongGeneration-v2-large
Captioner + judge	Gemini 3 Flash, native generateContent endpoint with a per-call audio-modality assertion