hakimi20274/xeus-l14-content-features

XEUS L14 speech/singing features and experimental content projections

Status / 状态

Research artifact, not a validated production SVC model. Intended initial visibility: private. 约 101.665 小时已选取音频对应的 XEUS L14 特征（包含留出数据）。不包含原始音频。 Features include eight languages. The v3 projection was fitted before French was added; French features in batch v6 were NOT used to fit the supplied v3 matrix.

Contents

• data/: lossless byte-planed NPY files in zstd level-6 compressed TAR shards.
• manifests/: one JSONL per batch, mapping every feature to its source audio clip.
• weights/run_v3/: previous evaluated experimental shared projection, a Chinese singing canonical-form W1 baseline, reference PCA and metadata.
• weights/run_v2/: prior version, retained for comparison.
• reports/: same-phoneme EER and per-layer tests, plus available evaluation records.
• scripts/: experimental extraction/fitting code snapshots. Several use /content paths and require adaptation; not a turnkey supported package.
• provenance/statistics.json: measured duration, clip and identity-label counts by batch/source/domain/split.
• shards.jsonl: generated shard contents and checksums. PACKING_COMPLETE.json is created only when every batch is packaged.

Feature format and exact restoration

Each restored .npy is [frames, 1024], little-endian float16. The encoder is espnet/xeus, checkpoint model/xeus_checkpoint_new.pth, encoder block 14 (1-based), without the encoder's terminal after_norm. Inputs are mono, resampled to 16 kHz. Extraction used FP16 autocast, one utterance at a time, eval mode. Audio longer than 10 seconds was split into 10-second chunks and the resulting features concatenated. There is no explicit per-frame timestamp or chunk-boundary array. Reconstruct chunk timing using the extraction implementation, not simply t*20ms across all chunks.

The original NPY header is preserved; the float16 payload is rearranged into two byte planes before archiving. This is lossless (no quantization beyond the stored FP16). Do not load .npy.bp directly with numpy. Restore with:

pip install numpy zstandard
python scripts/restore_features.py --input data --output restored

The restoration script verifies each original NPY SHA-256. This is distinct from manifest sha256, which hashes the original encoded audio bytes.

Source traceability

Each manifest row has dataset, language, domain, speaker, source, seconds, frames, feature, split, sha256; some also have text. feature is relative to the restoration directory; original_feature_path records the historical VM path. source identifies a dataset clip or archive member, NOT necessarily an offset in a complete original song. Some Common Voice paths are old cache paths; the clip basename and corresponding dataset/version identify the clip.

Batch overview:

1. GTSinger Japanese/Russian; LibriTTS-R English; AISHELL-3 Mandarin; MLS Polish.
2. M4Singer processed third-party mirror, OpenCpop excluded.
3. OpenSinger subset from CrawlSinger-OS archive (original source license retained).
4. Common Voice 18 Arabic and Zeroth-Korean.
5. Common Voice 17 Japanese/Russian.
6. Common Voice 17 French subset.

Identity labels are dataset namespaces, not globally deduplicated natural persons. OpenSinger mirror exposed 76 gender/ID combinations versus 66 singers reported in its original paper; do not advertise 76 verified distinct people. heldout is split by identity label before fitting, but cross-corpus real-person overlap has not been audited.

Weights

W_shared_experimental_L14_r256.npy has shape [1024,256], float32. Inference:

import numpy as np
x = np.load('restored/features_v1/example.npy').astype(np.float32)
w = np.load('weights/run_v3/W_shared_experimental_L14_r256.npy')
z = x @ w

No reference PCA or reference mean is applied at inference. This is NOT canonical USCF W1: it uses local whitened SCF, PCA-reference orthogonal alignment, and shared regression. The extra reference arrays document the fitting procedure. W1_Chinese_singing_baseline_L14_r256.npy follows the original-form W1 regression, but is only a small Chinese-singing baseline (up to 16 fitting singers), not a multilingual final W1. A matrix fitted on L14 must not be reused on other layers.

Evaluation and limits

The same-phoneme test used 240 English speech/singing clips from three GTSinger identity labels absent from fitting manifests. Official phoneme boundaries were used; references/queries were separated by song, with a separate rotating speaker for phoneme references. This is not a TIMIT replication and is a small-sample probe. Speaker EER near 50% means difficulty for this particular verifier, not absence of all recoverable identity. Phoneme EER must also be checked. Shared projections improved phoneme cosine discrimination in this test, but did not demonstrate clear additional speaker removal in singing. No downstream SVC listening evaluation has been conducted. Per-layer tests used the same small English pool, not a multilingual benchmark. See reports for exact scores and exclusions; French is not covered by those evaluations.

Provenance and reproducibility

Encoder code: wanchichen/espnet ssl, commit 5b52d57b4f872ff7babded35316a79642e9e6c12. USCF reference: HSTEHSTEHSTE/uscf, commit 784a97aeb2c1446bb285e55a56dc7513d967aa4b. Checkpoint compatibility adjustments are recorded in scripts: ignore disabled normalize_output=false; rename positional convolution checkpoint keys to older branch names; strict parameter loading. Equivalence to newer training implementation has not been separately established. Data source repositories were not all pinned to immutable revisions; source audio hashes assist matching exact clips.

Licensing

See LICENSE_NOTES.md. Do not treat this mixed collection or the experimental weights as MIT/CC0 or automatically commercially usable. Verify redistribution rights before making the repository public. A private upload does not itself resolve license questions.

v4 update (French included)

weights/run_v4/ adds French fit identities from batch v6; heldout identities remain excluded. v3 remains unchanged and excludes French. v4 is the same experimental shared-projection method, not canonical USCF. See run_v4/fit_metadata.json, quality_report.json and phoneme_report.json for measured results; completion is not a claim of production SVC quality.