laion/MOSS-v2-12to32-Enhancement-Tokens

MOSS v2 audio enhancement: 12 → 32 codebooks

Incremental release: new recordings are committed while encoding continues. Each commit contains complete clean/augmented pairs and an updated manifest. Pin a commit for reproducible use.

Published: 221,557 unique recordings, 805.44 hours (including sub-frame padding). Planned source corpus: 803.20 hours, all splits; 221,557 recordings.

Language	Original source hours
de	403.14
en	100.06
es	150.00
fr	150.00

Data and codec

There are separate ears_clean, ears_augmented, mls_hq_clean, mls_hq_augmented, hui_clean and hui_augmented configurations. A configuration appears when its first completed shard is uploaded. One row represents one original recording. Join clean and augmented rows by sample_id; every pair has the same clean32 target and split. clean12 is (frames,12), clean32 is (frames,32), and augmented12 contains three (frames,12) buffers. All buffers are little-endian uint16 codebook indices in [0,1023]. The codec runs at 48 kHz stereo, 12.5 frames/s (80 ms/frame). Mono is duplicated. Clean12 is the exact first-12-codebook prefix of a single clean32 encode. Codec: OpenMOSS-Team/MOSS-Audio-Tokenizer-v2; pinned revision f6e20e543b33d2c252a7ef71bdf8aa71e5ff9169. Encoder BF16, residual quantizer FP32, upstream public batch_encode with 0.96 s streaming chunks. End padding is recorded; no future source frames are added to model training inputs. Here 'clean' means the original supplied recording after normalization/resampling. EARS and MLS mirrors already use lossy Opus; 32 codebooks do not recover their pre-Opus masters. HUI uses 44.1 kHz PCM WAV and the mirror does not identify Full/Clean membership per recording.

Augmentations

Three distinct mild, deterministic recipes per original: bandwidth 8/12/16/24 kHz, white/pink/brown noise (26–32 dB SNR), gentle low/high-pass filtering and shelves, early reflections, short quiet reverb, mild saturation, small amplitude modulation, or a ±12-cent / 18%-wet global pitch-snapping effect. This pitch effect is a limited approximation, not a local AutoTune processor. There are 18 recipes. Every six source recordings use each recipe once; sample seeds, parameters, gain and original/padded lengths are in metadata_json. All variants retain the clean target's time alignment.

Splits

EARS retains the official speaker-disjoint train/validation/test splits. MLS and HUI use deterministic speaker-group hash splits (approximately 90/5/5% of groups). Variants remain together. Hours per split need not be 90/5/5%.

Load and decode

from datasets import load_dataset
import numpy as np
ds = load_dataset("laion/MOSS-v2-12to32-Enhancement-Tokens", "ears_clean", split="train", streaming=True)
row = next(iter(ds))
hq = np.frombuffer(row["clean32"], dtype="<u2").reshape(row["frames"], 32)
lq = np.frombuffer(row["clean12"], dtype="<u2").reshape(row["frames"], 12)
# codec.decode expects [codebooks,batch,frames]: hq.T[:,None,:]

For an augmented row, reshape each of its three augmented12 buffers identically. Source revision, shard SHA256, speaker, book/text where supplied, codec settings and augmentation provenance are retained.

Licenses and attribution

See LICENSES.md and release_licenses.json. Source-specific terms apply to clean and augmented token derivatives. EARS is noncommercial. The sources are not all relicensed under a common permissive license.

philgzl/ears

Julius Richter, Yi-Chiao Wu, Steven Krenn, Simon Welker, Bunlong Lay, Shinji Watanabe, Alexander Richard and Timo Gerkmann; EARS, Interspeech 2024; Opus mirror by philgzl.

Source revision: d16129286b1135851e6287d01f85d3a1fe532c3e. cc-by-nc-4.0. Pinned mirror.

philgzl/mls-hq-urgent-track1

Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve and Ronan Collobert; Multilingual LibriSpeech (MLS); MLS-HQ selection by the URGENT challenge, kohei0209; Opus mirror by philgzl.

Source revision: e32da0f8bda4807c9ea388bc1e5abbc1e6ede7ce. cc-by-4.0. Pinned mirror. The supplied mirror declares CC0-1.0. The original MLS resource declares CC-BY-4.0. This release retains the original attribution and records both declarations.

Paradoxia/opendata-iisys-hui

IISYS / Hof University researchers; HUI-Audio-Corpus-German, 2021; original LibriVox narrators retained in speaker/book metadata; Hugging Face mirror by Paradoxia.

Source revision: 44cc3a88870e4a64f142ad5bd808f87a5973f6a3. mit. Pinned mirror. MIT is the mirror's declared dataset license. Its card supplies no separate copyright notice or license text. The official generation-code repository is Apache-2.0; that software license is not substituted for the dataset declaration.

Follow-up: distortion and volume restoration

Published 36,000 base clips / 100.000 hours per task; planned 36,000 / 100.0 hours per task. Two configurations: distortion_100h_v1 and volume_100h_v1. Each row is exactly ten seconds / 125 frames. Both tasks share the same clean32 target for a given clip_id. input12 and clean32 are little-endian uint16 byte buffers, shapes (125,12) and (125,32). Targets are new codec encodings of clean source clips after one static gain: active speech RMS -20 dBFS with sample peak <=0.95. No enhancer predictions are used as training targets. Distortion: mild/medium saturation, controlled hard/asymmetric clipping, cubic waveshaping, 8/10-bit quantization and 8-bit mu-law quantization. Levels are approximately RMS matched for this task. Volume: -18/-30/-42 dB attenuation (2 dB jitter), smoothly varying quiet gain, or hard overload affecting 0.5/2/5/10% of samples. Overload is deliberately clipped and is not subsequently normalized away. Train split only. Duration-stratified across sources/languages and speakers; the original validation/test speakers are excluded. Short originals may be concatenated within the same source/language/speaker. No artificial silence or codec padding counts toward the ten seconds. Sampling: distribution. Planned language hours per task: {"de": 52.102777777777774, "en": 12.613888888888889, "es": 15.694444444444445, "fr": 19.58888888888889}. All source components, offsets, raw shard hashes/revisions, augmentation seeds/parameters, target gain, codec revision and actual encoding precision are retained in metadata_json. See follow-up manifest, recipe and provenance, and source licenses. The existing source-specific licenses continue to apply; EARS remains CC BY-NC 4.0.

Unseen wideband expansion v2 — 1000h target

Published: 64,918 base clips / 180.328 unique hours / 721.311 conditioning hours. Planned unique language hours: {"de": 250.0, "en": 250.0, "es": 250.0, "fr": 250.0}. Counts are targets until encoding and remote verification complete. Each ten-second clip has one shared clean32 codec target, its exact clean12 prefix, and three independent corrupted12 views. No model-generated targets. Twenty new duration-preserving recipes; old eighteen recipes remain available. The two existing distortion/volume task families are retained separately. MLS-HQ: unseen original utterances including exclusion of all their trimmed subsegments, and excluding original held-out LibriVox speakers across languages. English: NVIDIA HiFiTTS-2 44kHz metadata, train/single-speaker/low-error/high-bandwidth filters. No repeated audio to fill quotas. See complete recipe, storage and training documentation and licenses and attribution. New source audio remains CC BY 4.0. Existing EARS data remain CC BY-NC 4.0; the extension does not change the older source licenses.

Public data and completed group-relative RL pilot

This dataset is paired with laion/MOSS-v2-12to32-Enhancer-60M. The completed SAPO / GRPO-family model, runnable inference code and checkpoint reproduction instructions are on its rl-sapo-v1 branch.

Distinguish these three data types

1. Original supervised data: EARS, MLS and HUI; clean12 and three corrupted12 views share a clean32 codec target. Original train/validation/test splits and per-source licenses are retained.
2. Wideband expansion: expansion_v2/1000h/train/*.parquet; ten-second clips from unseen MLS-HQ and filtered NVIDIA HiFiTTS-2, each with clean12 plus three corrupted12 views and one shared clean32 target. The directory name is a planned 1,000 h target. Published counts in manifest.json are the actual verified data, and can be less than the target.
3. RL archives: rl_sapo_v1/rollouts/*.npz; fixed input12, clean32 target, eight model-generated candidate32 sequences, behavior probabilities, fixed-reference clean-label NLL, seeds, ranking and reward metadata. Candidates are not codec ground-truth labels. These archives support replay of the 104-group / 208-update pilot. The new wideband expansion was not used in that completed pilot.

All arrays use categorical indices 0–1023, with distinct meaning per codebook. MOSS v2 uses 48 kHz stereo at 12.5 frames/s. Ten seconds is 125 frames; the 32-book target rate is 400 indices/s. Twelve-book input rate is 150 indices/s.

Load newly published supervised rows

from datasets import load_dataset
import numpy as np

rows = load_dataset("laion/MOSS-v2-12to32-Enhancement-Tokens",
                    "unseen_1000h_v2", split="train", streaming=True)
row = next(iter(rows))
target = np.frombuffer(row["clean32"], dtype="<u2").reshape(125, 32)
clean = np.frombuffer(row["clean12"], dtype="<u2").reshape(125, 12)
augmented = np.frombuffer(row["augmented12"], dtype="<u2").reshape(3, 125, 12)
assert np.array_equal(clean, target[:, :12])
print(row["language"], row["recipes"], row["license"])

The three augmentation recipes are distinct per clip and balanced within language. Twenty new duration-preserving recipes supplement the original eighteen families, with the two earlier distortion and volume restoration tasks retained separately. Missing future data are not silently repeated to fill quotas. Every published row preserves source, speaker, deterministic seeds, augmentation parameters, original frame/window identity and codec pin. Use the manifest's immutable shard SHA256 values when downloading.

Inspect an RL group

import json, numpy as np
from huggingface_hub import hf_hub_download, HfApi
repo = "laion/MOSS-v2-12to32-Enhancement-Tokens"
commit = HfApi().repo_info(repo, repo_type="dataset").sha
files = HfApi().list_repo_files(repo, repo_type="dataset", revision=commit)
name = sorted(f for f in files if f.startswith("rl_sapo_v1/rollouts/")
              and f.endswith(".npz"))[0]
path = hf_hub_download(repo, name, repo_type="dataset", revision=commit)
with np.load(path, allow_pickle=False) as group:
    metadata = json.loads(str(group["metadata"]))
    winner = group["candidates32"][metadata["ranking"][0]]
    assert group["input12"].shape == (125, 12)
    assert group["clean32"].shape == (125, 32)
    assert group["candidates32"].shape == (8, 125, 32)

Validation archives live under rl_sapo_v1/evaluation/; do not use them for training. Full composite Best-of-8 rankings use clean targets and therefore are not a target-free deployment procedure. The model repository's replay script consumes these exact archived groups and compares reconstructed adapter tensors against the published checkpoints.

Reproducibility and licenses

Pin HF commits for both repositories. The model's release_manifest.json pins the completed RL archive. New expansion counts remain incremental and are tracked separately; they do not alter the completed pilot's input set. English documentation, code, contracts and license files are backed up under rl_sapo_v1/ and expansion_v2/1000h/ as well as in the model repository.

EARS is CC BY-NC 4.0, MLS CC BY 4.0, and HUI retains the mirror's declared MIT condition. New MLS-HQ/HiFiTTS-2 expansion rows are CC BY 4.0. Keep each row's license and attribution. See the repository's LICENSES.md and the expansion's LICENSES.md. Code is MIT; tokenization does not change source licenses. The new supervised mixture includes all recorded input views, with shared targets stored only once.

Two-epoch supervised continuation

Two-epoch supervised continuation with all recorded augmentation views

Model branch: allviews-two-epoch-v1. Dataset: laion/MOSS-v2-12to32-Enhancement-Tokens.

This run starts from the selected SAPO group-52 merged checkpoint, chosen by the highest fixed held-out greedy composite reward among the supervised base, group 52 and group 104. This is the best observed checkpoint under that small comparison, not proof of universal superiority. Merged BF16 weights are promoted to FP32 master weights for full-parameter training; the original unmerged adapter policy remains separately available on rl-sapo-v1.

Every original and newly encoded recording contributes its clean12 view and all three recorded corrupted12 views once per epoch. The previous 100 h distortion and 100 h volume restoration tasks each contribute their one explicit input view. The duplicated 400 h rehearsal copy is excluded. Original validation/test splits remain held out. Two independently shuffled epochs use one continuous optimizer and LR schedule, not two optimizer resets.

The initial frozen available-data mixture contains 180.328 h of new unique audio; the planned 1,000 h expansion is not represented as completed. It contains 173,960,816 conditioning frames per epoch: 5,566,746,112 target indices per epoch, 11,133,492,224 across two epochs, and 679,535 optimizer steps at effective batch 512. Earlier throughput suggests roughly eight days on the eight host GPUs; use the new run's measured ETA once available. Actual quantities, source manifests and view counts are in training/preparation.json and the dataset's supervised_continuation_v1/manifest.json. The frozen run does not silently acquire later uploads. Clean32 targets are codec ground truth; RL-generated candidates are not substituted as supervised labels.

Training schedule and resources

• All 60,858,112 parameters are trained using fresh AdamW.
• Two epochs; all recorded views in each epoch.
• Peak LR 7.5e-5; 5% warm-up over the entire two-epoch run, cosine to 7.5e-6.
• AdamW betas 0.9/0.95, weight decay 0.01, gradient clipping 1.0.
• Eight GPUs, microbatch eight per rank, eight accumulation steps, effective batch 512 frames.
• BF16 autocast, FP32 master parameters/optimizer, gradient checkpointing.
• Per-process GPU cap 2.25 GiB, preserving 1 GiB physical free headroom.
• 32 CPU cores; two loader workers per rank; mmap input shards.
• Checkpoint every 1% of the complete run, with fixed teacher-forced validation.
• Every checkpoint, optimizer state and exported inference model is pushed to HF.
• Keep the best plus latest two checkpoints locally; older continuation checkpoints are deleted only after remote size/SHA256 checks succeed.
• Physical free-disk reserve: 30 GB; task persistent budget: 80 GB.

Teacher-forced cross entropy is the checkpoint selection metric for this continuation. It does not replace perceptual evaluation. The old completed RL release and its checkpoints are preserved. Existing unrelated processes are never terminated by the launcher.

Inference

Install the package from this branch using python -m pip install -e ..

python -m moss_enhancer.inference \
  --model laion/MOSS-v2-12to32-Enhancer-60M \
  --revision allviews-two-epoch-v1 --tokens input12.npy \
  --output output32.npy --device cuda:0 --wav output.wav

This loads the selected full enhancer directly; attaching the old RL adapter again would apply it twice.

Reproduce or resume

On the original host, the detached publication/training supervisor is started with python3 scripts/start_publication_and_training.py. It publishes the completed RL release, backs up the frozen mmap dataset, checks GPU headroom, then launches training and a continuous verified checkpoint uploader. It defaults to the already encoded and remotely verified 180.328 h subset. --require-complete-expansion rejects this subset rather than silently claiming that the full planned 1,000 h exists.

Pin the model commit. checkpoint_manifest.json lists the exact dataset commit, checkpoints and hashes. Download supervised_continuation_v1/** from that dataset revision. Its portable manifest.json references the downloaded mmap/*.npy files relative to itself. Source-specific licenses and original train/validation/test assignments are retained.

Set store_manifest in training/training_config.json to that manifest, initial_checkpoint to the downloaded initial/training_state.pt, and output_dir to a new local directory. All other hyperparameters and the initial SHA remain fixed. Then run:

python -m torch.distributed.run --standalone --nproc_per_node=8 \
  training/train_distributed.py --config training/training_config.json

To resume, put a published training_state.pt in the output directory as step-XXXXXXXX.pt. The trainer validates architecture, dataset-manifest checksum, epoch count, initialization identity and optimizer schedule. The trainer uses the same portable manifest bytes that are published with the dataset. Relocate the enclosing dataset directory without editing that manifest; changing only the config's file paths preserves its checksum. Numerical identity across GPU/driver/kernel versions is not guaranteed.

Code is MIT. The model remains CC BY-NC 4.0 because the base includes EARS. Keep source-specific token licenses and attribution from the dataset repository.