The dataset viewer is not available for this split.
Error code: RowsPostProcessingError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Synthetic Singing Voice — Notes, MIDI and Acoustic Tokens
A synthetic solo-singing corpus with original audio and metadata, automatic note-event dictionaries, standard MIDI, complete flattened MIDI and existing MOSS-v2 acoustic codes. All source recordings are synthetic.
Status: Complete: all 206,633 recordings have note annotations.
Processing code · Annotation/listening laboratory · Experimental singing checkpoint · Annotation format and reproduction
Inventory
| Source group | Available original recordings | Audio hours |
|---|---|---|
| Lyric singing in five languages | 49,993 | included below |
| Nonlexical singing on syllables | 10,000 | included below |
| Both newer synthetic-voice groups | 59,993 | 328.859 |
| LeVo solo singing | 146,640 | 1174.921 |
| Total | 206,633 | 1503.780 |
The planned 50,000 + 10,000 newer songs yielded 49,993 + 10,000 available raw recordings. Of these, 49,881 + 9,988 had complete observed scripts and were used in the earlier checkpoint. The additional 124 raw recordings are retained here for note annotation with their missing-script flags. Seven missing generations are not invented or duplicated. The historical “Singing Voice 1K” name is retained by its source repository; the measured durations above describe this exact release inventory.
| Language of newer lyric songs | Recordings |
|---|---|
| English | 9,997 |
| German | 10,000 |
| Dutch | 10,000 |
| French | 10,000 |
| Spanish | 9,996 |
There are another 10,000 nonlexical clips labelled zxx. The LeVo group has
intended English lyrics; ASR can misrecognize the words or language. No human
language-verification claim is made. Original genre, mood, register and synthetic
speaker-alias fields are retained where available. LeVo singer identities are
unknown; no shared identity is invented for those recordings. Requested genres
and emotions describe intent, not verified acoustic outcomes.
Annotation and training fields
Default: ROSVOT + Procedural, smoothed pitch splits and gap filling
(rosvot_procedural_segmented), the frozen engineering default from the lab.
The original vocals are used directly: no extra stem separation changes their
duration or existing token alignment. Original audio and q12/q32 array bytes
are unchanged. Both uncombined note baselines and continuous F0 are retained.
Each recording dictionary contains the original metadata, source IDs/hashes,
prompt_original, general, midi_annotation, script, prompt_with_midi,
language/modality/voice labels and explicit review flags. The new text layout is:
GENERAL: original overall singing directions
MIDI Annotation:
complete flattened MIDI token stream
SCRIPT:
original observed words and inline performance directions
The old GENERAL/SCRIPT prompt is preserved separately. Standard MIDI is .mid;
the float-time note dictionary is .notes.json; complete SVAMIDI tokens are
.midi.txt. They preserve the full timeline, including trailing rests, at a
10-ms serialization grid. The note JSON keeps original floating-point times.
Acoustic tokens are different from MIDI tokens. MOSS-v2 runs at 12.5 frames/s
(80 ms/frame). Twelve codebook indices per frame means 150 indices/s; 32 means
400 indices/s. Arrays have shape [frames, codebooks]. Neither “12” nor “32”
is a frame rate. The experimental model was trained with q12.
Of the LeVo clips, 1,696 lacked a usable observed training script in the
earlier preparation. These recordings are retained for note annotation, with
script_available=false and no invented observed words. The earlier exact
checkpoint training selection used 144,944 LeVo clips.
There are also 124 newer synthetic-voice recordings without a complete
observed script: 1,820 such records in the full raw corpus. Their intended
lyrics/directions remain in the original metadata but are not substituted for
a missing observed transcript.
Download and layout
hf download laion/synthetic-singing-voice --repo-type dataset --local-dir singing-data
data/synthetic_voices/*.tar and data/levo/*.tar contain self-contained audio,
q12/q32 codes, original metadata and the new note annotations. Matching compact
note_annotations/ TARs contain annotation-only sidecars. metadata/archives.json
lists committed archives and hashes; metadata/statistics.json gives exact
release counts. metadata/record_index.jsonl.gz provides lightweight per-record
joins, and the count CSVs summarize language, genre, mood and synthetic voice
aliases. The index describes the full inventory; during upload, consult
archives.json for currently committed shards. Seven pilot examples live under pilot/ and are not extra
independent evaluation data.
training_reproduction/ is the separate exact 392,256-example mixture used for
the experimental checkpoint: original GENERAL/SCRIPT instructions, q12 target
codes, reference bank and the complete three-epoch presentation schedule. It
includes the earlier non-singing/replay and other musical examples as well as
the selected songs. Training-data export/upload complete: True.
The later MIDI field was not used to train the published checkpoint.
Earlier sentence ASR/EARS annotations remain in the original datasets, joined by source ID and saved annotation locators; the newer synthetic-voice TARs also retain their source acoustic annotation member. ASR words, timestamps, acoustic scores and note estimates are automatic, not verified lyrics or ground truth.
Original repositories and licence
The original recordings and metadata remain in
the voice-acting dataset and
Singing Voice 1K. New compact note
sidecars are added there under singing_note_annotations_20261117/, joined by
source ID. Producer-local paths and operational fields are omitted in public
metadata; IDs and audio checksums retain the joins. The full internal records
are retained locally.
CC BY 4.0. Attribute LAION and the linked source datasets. Changes include automatic ROSVOT/Procedural note estimation, F0 extraction, MIDI serialization and additional prompt fields; original source audio/token bytes are unchanged. Processing code and upstream model/runtime licences remain separate.
Quality and limitations
This is raw, not fully filtered or fully human-annotated data. Further quality filtering and annotation are required. Singing, words, speaker identity, solo isolation, ASR timing, emotional labels and note predictions can be wrong. Wrong notes can increase coverage; RMVPE agreement is a shared-model proxy, not independent musical accuracy. Empty note sequences are retained when a method detects no pitched notes; successful serialization is not evidence of quality. The data are not an independent held-out benchmark for the released checkpoint.
- Downloads last month
- 114