Work in progress. This repository initially contains documentation only. Audio is not yet released. The selection has completed; fresh annotation, verification, reference matching and packed export are being prepared/running on JUPITER. Do not interpret the selected hours below as already uploaded hours. Payload publication additionally requires the source redistribution checks to be completed.

This collection is designed for research on emotionally diverse, caption-controlled speech synthesis and related audio understanding. It selects approximately 500 hours each of English, German, French, Dutch and Spanish real spoken audio, retaining existing MOSS Audio Tokenizer v2 targets and complete source metadata, and adds fresh Humanness Voice Annotator predictions. It is not newly synthesized Gemini speech, and this release does not claim a new trained TTS model.

Selected material versus released material

Language	Selected clips	Selected hours	Source scope
English	130,387	499.999811	Existing public RealSpeech SFT, only `role=sft`
German	132,576	499.999436	Existing public RealSpeech SFT, only `role=sft`
French	117,052	499.999186	285.649722 h podcasts +214.349464 h enhanced real MLS
Dutch	118,780	499.999744	Existing cut real podcast clips
Spanish	116,519	499.999978	Existing cut real podcast clips
Total	615,314	2,499.998155	Unique target UIDs

The approximately 6.64-second difference from exactly 2,500 hours avoids cutting words or padding with duplicates. EN/DE come from laion/tts-realspeech-sft-en-de. Other source families correspond to the locally available annotated product of podcast recordings and Multilingual LibriSpeech, including an enhanced MLS derivative. Audio enhancement/codec processing does not mean the original spoken performance was TTS-generated. Source-specific provenance and original metadata remain attached.

The initial French podcast-only pool supplied only423.64 eligible hours. Its final selection was jointly recomputed with already annotated French MLS; it was not silently padded with weak-emotion clips. No EuroSpeech, Common Voice, ElevenLabs or new Gemini generation was added to these real-speech targets.

How the selection works

The source RealSpeech recipe normalizes emotion, genuineness and vocal-burst blending before combining them:

selection utility = normalized genuineness + normalized blend + 2 × normalized emotion strength.

Emotion normalization uses the original pooled132.8-million-utterance mid-rank CDF. Genuineness/blend use source-by-language percentiles; published EN/DE percentiles are retained. Each assigned emotion must be in the true top three normalized emotions, have percentile≥0.90 and raw score>1. A stable full sort avoids the historical top-three partition/tie issue. No selected target UID or normalized transcript is repeated within its language.

Structural requirements are3–30 seconds, at least five words, successful source alignment, at least37 codec frames and all12 codebooks. Legacy speech/overall/background quality heads must be finite and outside their language pool's bottom decile. These are relative tail exclusions, not a human MOS or a universal “excellent quality” certificate: the legacy quality heads are not calibrated like today's EARS scores.

The first allocation is12.5 hours per emotion per language, filling scarce buckets first. Remaining time is explicitly redistributed only to other eligible emotions. All40 buckets occur in every selected language, but French and Dutch are not uniformly balanced because rare emotions lack supply. Fifteen explicit VoiceNet S_* styles are softly balanced within emotion queues: ASMR, authoritative, cartoon, casual, conversational, dramatic, formal, monologue, narration, newscast, playful, ranting, storytelling, technical and whispering. S_MONO is monologue, not monotone. The57 continuous VoiceNet dimensions also include properties that are not talking styles.

This is a selection utility, not an RL objective. Classifier predictions and inferred emotions are not human truth or psychological diagnoses.

What fresh annotation adds

Each original recording passes through:

Parakeet TDT0.6B v3 ASR and word timestamps.
Humaneness EARS Medium global/local emotion, VoiceNet, quality, event proposals and128-dimensional timbre /250-dimensional identity embeddings.
MOSS-Audio-8B voice-captioner for a full-recording English acoustic/performance description.
MOSS vocal-burst-captioner, given the full recording and both endpoints for each selected event.
Deterministic GENERAL/SCRIPT composition and independent word/event/timing-format verification.

The two inference environments remain loaded per GPU: Transformers5.14.1 for acoustics and4.57.1 for MOSS. The run uses256 GH200 GPUs on64 JUPITER nodes in resumable allocations. It reuses existing codec targets for unchanged original audio; it does not re-encode everything unnecessarily. Raw older teacher annotations remain separate from the new EARS/MOSS predictions.

Pinned component revisions:

Component	Repository / revision
Humaneness EARS Medium	`laion/humaneness-ears-base-medium` · `818506970d93c809a3295a0a02c40dee8ff2bfc3`
Parakeet v3	`nvidia/parakeet-tdt-0.6b-v3` · `541d1f99c6b0c3cd0b11a95167540bb8edefd82b`
MOSS voice captioner	`laion/moss-voice-captioner` · `86d3c6639fe24e10ed1feafb05fb2f9d3311dadd`
MOSS burst captioner	`laion/moss-vocal-burst-captioner` · `e6e7fae66a9f97a542a7f6df43239a18fe33bc63`
Procedural source	`LAION-AI/procedural-voice-captions` · `df9514b05f234c692a9e88df7633f5460346d2c9`

Prompt format

Illustrative only, not a claimed recording:

GENERAL: One close-miked speaker talks softly and hesitantly, then sounds relieved. The background is quiet.
SCRIPT:
(soft, hesitant; 1.20 seconds duration) I... wasn't sure.
[0.30 seconds pause]
("A short, unvoiced inhalation through the mouth.", 0.40 seconds)
(warmer, relieved; 2.30 seconds duration) ("A soft, breathy laugh.", 0.50 seconds) But I'm glad you came.

The total elapsed budget is4.20 s, not4.70 s: the inline laugh is already included in its enclosing speech segment. Standalone bursts and explicit pauses advance time; inline bursts do not advance it again. JSON retains exact starts/ends and overlaps. Segment durations can cover pause-split sentence pieces, with a retained source-sentence index. ASR gaps are not necessarily acoustic silence.

Descriptions and cue wording are English; recognized speech stays in the audio language. Both raw ASR text and aligned word text remain available. Older Humaneness Voice checkpoints may require the documented legacy adapter from (cue; N seconds duration) to (cue) [N seconds duration]; do not silently discard durations or read instructions as speech.

Packed files and codec details

Targets will be published as data/part-*.tar, with at most512 records per sealed TAR. For each UID:

Member	Meaning
`UID.mp3`	Original MP3 unchanged, or192-kbit/s mono playback derivative when source is FLAC
`UID.original.flac`	Included when needed to preserve the exact source audio used for the existing codec target and fresh annotation
`UID.moss.npy`	Existing integer `[frames,12]` MOSS Audio Tokenizer v2 target, values0–1023
`UID.json`	New canonical annotation, all retained source metadata, selection provenance and payload hashes
`UID.vclap.npy`	Original auxiliary embedding where supplied by the source product

The codec has12 codebooks per80-ms frame:12.5 frames/s and150 codebook indices/s. It is not12 indices/s and this release does not claim all32 codec codebooks. The original audio associated with the codec is identified explicitly. A lossy MP3 playback derivative is not falsely claimed to be the exact waveform from which old tokens were encoded.

See SCHEMA.md and loader.py. Reference assignments and their independently packed audio/codec crops will be under references/. No millions of loose audio or JSON files are required.

from huggingface_hub import hf_hub_download
from loader import iter_records

# Access requires a logged-in account with an approved gate request.
path = hf_hub_download(
    "laion/humaneness-realspeech-2500h-five-languages",
    "data/part-000-000.tar", repo_type="dataset", token=True,
)
for metadata, codes, mp3 in iter_records(path):
    print(metadata["uid"], codes.shape, metadata["annotation"]["prompt"])

Reference-audio policy

Good references should use a different utterance, contain5–15 seconds of actual audio, have good recording quality, and support rather than merely resemble the desired speaker. New vectors are measured on the actual codec-frame-aligned crop (up to14.88 s, with margin for encoding), not on a longer file that is later blindly truncated.

The historical Orange128 release used cosine≥0.94. Its structural-speaker-group audit did not establish a calibrated same-speaker probability; higher thresholds were not monotonically better on that proxy. New EARS-distilled128/250-dimensional vectors are searched in their own encoder space, not mixed into the old index. The new search prefers matching source speaker groups, requires distinct UID/transcript and adequate quality, and checks both timbre and identity vectors. All thresholds remain explicit heuristics. A nearest neighbor is never labeled a verified real-world identity. Poor matches remain unassigned.

Reference MP3s and source codec-frame slices are packed separately with a catalog and per-target status. Since MP3 encoding is lossy, metadata explicitly distinguishes the measured source crop and reused codec slice from the MP3 playback derivative. It remains possible to train without reference conditioning for unassigned examples.

Limitations, evaluation and intended use

Format verification establishes mechanical consistency, not correct ASR, emotion labels or event timing. ASR comparison against the older transcript is not comparison against human gold. NL/ES are supported by the ASR model but absent from the annotator's80,265-clip local-cue z-score reference, so multilingual cue calibration needs particular care. Raw scores, alignment decisions and model revisions are retained to support correction.

Suggested downstream evaluation separates languages, source speakers/sessions and overlapping episode segments; checks WER, genuineness, Content Enjoyment/production quality, emotion/style following, burst realization/blending and sentence-duration error; and includes human listening. The source selection is not a ready-made leakage-free benchmark split. Obtain appropriate rights and consent for downstream voice applications. Avoid impersonation, deceptive identity claims, harassment, surveillance and attempts to identify private speakers.

Licence and access

The requested licence for new annotations and the new collection is CC BY-NC4.0, attributed to LAION; exact confirmation is recorded before payload publication. Original recordings, original annotations, upstream model/code and pre-existing datasets retain their own licences and attribution. A new collection label does not relicense CC-BY material or grant rights that are absent upstream. Access approval is not a substitute for copyright, privacy or consent requirements.

Podcast redistribution rights are being checked before any related audio/transcripts/tokens are published. Until that gate is cleared, this is a documentation-only work-in-progress repository. We do not assert that possession of a local podcast file or token stream permits redistributing it under a new licence. Source owners may contact LAION through the repository discussion to raise provenance or rights concerns.

Downloads last month: 23

Total file size:

19.2 kB