laion/moss-voice-identity-repairs

MOSS voice-acting v2 -- repaired takes

For each voice profile, every take whose ECAPA speaker similarity to the voice's reference fell below 0.40, regenerated with that voice's identity LoRA (see laion/moss-voice-identity-loras) merged at scale 1.0 on top of the identical condition adapters at the identical lambdas.

Nothing here replaces anything. The original takes are untouched and remain part of the corpus; low-similarity takes are kept deliberately, because they are useful for training without reference audio, which is a different and legitimate use.

Upload complete (2026-09-09). All 499 voice directories are here: 30,058 tars (1,290.88 GB), 499 annotations.parquet (53.00 GB) and 499 report.json — 8,177,013 repaired takes. Every file was verified byte-exact against the cluster copy (repo_info(files_metadata=True) size == local size, 30,784 of 30,784 for the 489 voices delivered in the second pass; the earlier 10 matched by tar count). k325_age3_bg1, which the first pass had left with 48 of 55 tars and no sidecars, is now complete.

---

Layout

One directory per voice profile.

<voice>/part-<NN>-<NNNN>.tar     WebDataset shards: the audio
<voice>/annotations.parquet      one row per repaired take, 125 columns
<voice>/report.json              that voice's before/after and cost report

<NN> is the generation arm and <NNNN> the shard within it; both are storage detail and carry no meaning you need. A voice's tars are its complete repaired output.

Where the audio is, and in what format

The audio is in the tars, and nowhere else. Two members per take, sharing one stem:

member	format	detail
<wds_key>.mp3	MPEG-1 Layer III, 48 kHz, 160 kbps CBR, 2-channel	the repaired take, exactly the bytes the generator wrote — not re-encoded
<wds_key>.json	JSON	the take's full generation record

The MP3s are two-channel files carrying duplicated mono. That is how the generator writes them. Read back from the frame headers of the shipped bytes, not inferred. It matters: a decoder that flattens stereo frames instead of taking one channel returns a signal exactly twice as long, at half speed. Every mainstream decoder (soundfile, librosa, torchaudio, ffmpeg) does the right thing; a hand-rolled frame reader may not.

There are no codec tokens here

This repository ships no .moss.npy and no .vclap.npy — unlike laion/laion-voice-profiles-annotated, whose tars carry four members per utterance. Here it is two: audio and JSON. If you need MOSS codes for these takes, they are in that repository's origin=repair partition, under the same key (see the join below), or you can encode them yourself from the MP3.

What this repository does ship instead of tokens are embeddings, inside the parquet: src_spk_emb / rep_spk_emb / org_spk_emb (ECAPA), rep_tbr_emb / org_tbr_emb (WavLM timbre) and src_voiceclap, all as raw float16 buffers.

The join

wds_key is the tar member stem. The parquet carries no shard column, so:

import tarfile, glob, io, json, pyarrow.parquet as pq

ann = pq.read_table("anime_000/annotations.parquet",
                    columns=["wds_key", "rep_spk_sim", "org_spk_sim", "text"])
want = {str(k): i for i, k in enumerate(ann["wds_key"])}

for path in sorted(glob.glob("anime_000/part-*.tar")):
    with tarfile.open(path) as tf:
        for m in tf:
            stem, _, ext = m.name.rpartition(".")
            if ext == "mp3" and stem in want:
                audio = tf.extractfile(m).read()
                ...

One voice is a few dozen tars, so a single streaming pass is the cheap way to do this.

If you want the tar name without opening anything, take it from the annotation corpus: in laion/laion-voice-profiles-annotated partition origin=repair, uid equals this repository's wds_key and src_shard names the tar in this repository. Verified on anime_000: 3,865 of its 3,872 rows join, and every src_shard they carry is a file that exists here.

audio_key is not the member stem and is not unique across runs; run_dir is stored beside it and every join that uses audio_key must use both. Prefer wds_key.

Columns

125 columns, one row per repaired take, identical across every voice (checked on four). The full generated reference is in COLUMNS.md.

Every row carries three groups of columns, and none of them is derived from another:

group	columns	what it is
src_*	63	the corpus's own stored annotation of the original take, complete: every score component (strength_raw, w_blend, z_containment, contained, dur, wer, spk_sim, the z-terms, the multiplicative factors), so any later re-ranking needs zero regeneration
org_*	21	the original audio re-scored here through the same sensor stack as the repair (ECAPA, WavLM-tbr, Whisper-large-v3-turbo, the corpus's fast scorer)
rep_*	21	the repaired take, measured identically

Stored WER came from Parakeet and is not comparable to Whisper WER, which is why the before side is re-measured rather than inherited. Compare rep_* against org_*, not against src_*.

The remaining 20 columns are identity (voice, wds_key, audio_key, gid, cand, subset, block, lang, is_holdout, run_dir), content (text, caption), the adapter stack (adapters), the repair bookkeeping (repair_adapter, repair_adapter_sha, repair_scale, repair_epoch, repair_arm, floor) and orig_rescored.

repair_adapter_sha means every take records exactly which weights produced it.

report.json

One per voice, written by the build. It holds the LoRA's training curve and chosen epoch, the candidate-pool statistics, and the before/after means over all of that voice's repaired takes — spk_sim, tbr_sim, share above the 0.40 floor, and the cost side (strength_raw, blend, genuineness, wer, quality, reward, emo_strength) so the trade the repair makes is visible without opening the parquet.

Scale, and what the repair achieved — numbers from the 499 report.json

voices	499 (of the corpus's 500; one had no takes below the floor)
repaired takes	8,177,013
WebDataset shards	30,058, from 16 to 72 per voice
audio	1,290.88 GB · annotations.parquet 53.00 GB
GPU-hours spent	10,765 (21.6 per voice)
ECAPA similarity to the reference, org → rep	0.266 → 0.487
share of takes above the 0.40 floor, org → rep	0.2 % → 72.5 %
gain in percentage points, per voice	mean 72.3, median 76.8, min 14.2, max 96.1

Three things that must travel with any result from this data

1. The 0.40 floor is this project's ECAPA scale, not an established perceptual fact. The independent WavLM-tbr embedder (Orange/Speaker-wavLM-tbr, published same-speaker threshold 0.472) puts 75.4 % of the takes selected as "failures" above its own threshold. Both embedders are in every row (*_spk_sim, *_tbr_sim), so the selection can be redone.
2. Identity is bought, not free. Averaged over the 499 voices the repair cost −0.338 target-emotion strength, −0.061 blend, −0.069 genuineness and +0.252 WER for its identity gain. Each voice's report.json → cost has its own numbers.
3. A before/after delta is not the adapter's effect on its own. Production regenerates one arm, so there is no per-voice control. The reference study on emolia_c1699 (the worst-affected voice, so an upper bound) found that merely re-rolling these takes with no voice adapter lifts them from 0.2 % to 14.0 % above the floor — they are the unlucky tail of their groups. The adapter took the same population to 43.7 %: roughly 13.8 pp is regression to the mean, roughly 29.7 pp is the adapter.

Upload provenance

Second pass 2026-09-09 11:26–12:30 Berlin from a JUPITER login node via HfApi.upload_large_folder (16 workers, Xet), sustained 327–340 MB/s measured as committed bytes per wall-clock second over 5-minute windows. Files land in batches of 256 across voices, so during an upload a voice directory can be missing any of its three file kinds — that says nothing about the data; count *.tar.done on the cluster if you need the state mid-flight.