README.md

---
license: cc-by-4.0
task_categories: [text-to-speech, audio-classification]
language: [en, de, fr, es, it, nl, pl, pt]
tags: [speech, emotion, voice, trajectory, tts, moss, webdataset]
size_categories: [100K<n<1M]
configs:
  - config_name: default
    data_files:
      - split: train
        path: traj-t80-*.parquet
---

# LAION Emotional-Trajectory Speech — tier T≥0.80

**319,765 crossfaded speech trajectories · 4,482 audio-hours · 1,598,825 source clips**

A *trajectory* is a short sequence of 5 consecutive utterances **by one
speaker** whose measured emotion or voice character moves monotonically from one end of
the corpus distribution to the other. The clips are joined into one continuous audio file
with equal-power crossfades, the joined audio is **re-tokenized** with MOSS-Audio-
Tokenizer-v2, and every chain carries a two-level caption.

This is the **T≥0.80** rung: every chain crosses at least **80 % of the corpus range**
on its named dimension while never stepping more than 25 % between adjacent clips.

## Why a trajectory dataset

Most expressive-TTS corpora label a clip with *one* emotion. That teaches a model to hit
a target but not to *travel* — to start guarded and end furious, to slide from amusement
into contempt. These chains are examples of the travel itself, with the change measured
rather than asserted.

## The tier ladder — strictly nested

| tier | chains | hours | en % | de % | other % |
|---|--:|--:|--:|--:|--:|
| T≥0.20 | 4,381,192 | 44,999 | 62.8 | 21.4 | 15.8 |
| T≥0.25 | 2,881,179 | 33,590 | 62.2 | 26.5 | 11.3 |
| T≥0.40 | 1,594,365 | 19,603 | 61.9 | 34.2 | 3.9 |
| T≥0.50 | 1,007,755 | 13,371 | 61.0 | 37.8 | 1.2 |
| T≥0.60 | 783,747 | 10,423 | 60.6 | 39.2 | 0.3 |
| T≥0.70 | 615,517 | 8,194 | 59.7 | 40.3 | 0.0 |
| T≥0.80 **(this release)** | 319,765 | 4,531 | 59.6 | 40.4 | 0.0 |

Higher T is a **strict subset** of every lower T (verified by set containment, all four
rules × all seven rungs). So you can train on T≥0.20 and evaluate on T≥0.80 knowing the
harder set is contained in the easier one, and a curriculum needs no re-download.

## How a chain qualifies

Scores are mapped to a **tie-aware mid-rank ECDF** over all 132,833,726 annotated
utterances, so "moved 0.25" means "crossed 25 % of the whole corpus" identically on every
dimension. Writing `u_e(i)` for that percentile of dimension `e` at clip `i`, with
`A = argmax_e u_e(first)` and `B = argmax_e u_e(last)`:

| rule | family | requires |
|---|---|---|
| `B1` | `emotion` | one-sided: `\|Δu_B\| ≥ T`, per-step ≤ C on axis B |
| `AB2` | `emotion_twosided` | both named axes move by ≥ T, per-step ≤ C on both |
| `PXR` | `proxy_spearman` | as AB2, but an endpoint that cannot ramp certifies smoothness on a correlated proxy axis |
| `VN1` | `voicenet` | one of 57 VoiceNet voice descriptors sweeps by ≥ T |

`C = 0.25` throughout. A chain may satisfy several rules; the `rules` column lists all of
them and the row appears **once**.

> **PXR caveat.** Only the *Spearman* proxy family is present in the source chain table.
> A tail-lift proxy family exists upstream (it covers 1,553 of 1,560 ordered emotion pairs
> against Spearman's 1,475) but has no rows here, so every `PXR` chain in this release is
> Spearman-derived and labelled `proxy_map = "spearman"`.

## The speaker filter — strict, and no voice conversion

Mined chains must satisfy **both** conditions on `Orange/Speaker-wavLM` cosines:

```
min_cos_consec  >= 0.80     every adjacent pair
min_cos_anchor  >= 0.80     every clip against the first
```

Neighbour-only similarity does not chain — A can resemble B and B resemble C while A and
C are plainly different people — so the anchored condition is what catches drift, and
requiring both is the strict reading. **Chains with no measurement are dropped, not
kept.** No voice conversion is applied anywhere: a chain is one real speaker or it is not
in the dataset.

`vprof_vc` chains are exempt from this filter **and only from this filter**: one voice
profile is one cloned voice by construction, so there is no identity to verify. They
carry `cos_source = "vprof-cloned-voice"`.

## Rendering

* Every clip levelled to **exactly −20.0 dBFS RMS** before joining. MOSS's own loudness
  normaliser targets −20 dBFS but clamps gain to ±3 dB, and 59 % of corpus clips hit that
  clamp; pre-scaling makes its gain zero, so the seam step is zero *by construction*.
* **150 ms equal-power (cos/sin) crossfade** at every join, shortened to 100 ms on a hot
  onset or tail, never more than 25 % of either segment. A linear crossfade sums to less
  than unit energy at its midpoint and dips audibly.
* Peak guard applied **once** to the finished chain.
* 48 kHz mono, MP3 96 kbps CBR.

## Re-tokenization — and why it was necessary

The crossfaded concatenation is **new audio**. The constituent clips' existing MOSS codes
describe the clips, not the chain, and MOSS's codec blocks are causal — decoding clip B
with clip A in context changes B by about −7.5 dB relative error over its whole length.
So the chain is re-encoded from the rendered waveform:

`OpenMOSS-Team/MOSS-Audio-Tokenizer-v2`, **12 codebooks × 1024, 12.5 fps**. One frame is
12 tokens = 80 ms. (The codec ships 32 quantizers; the model consumes 12. 32 is depth,
not a block size.)

Stored as `uint16 [T, 12]` in `<chain>.moss.npy`, with `moss_frames == floor(dur_s ×
12.5)`.

## Captions — two levels, both shipped

### Inline, per segment
A screenplay: `(emotions · voice descriptors) the words spoken`, one line per clip.
Descriptors identical across the whole chain are hoisted out into
`constant_descriptors`, because a value that never changes says nothing about a
trajectory.

### General, for the whole concatenation
Derived by **scoring the rendered audio**, not the source clips. Two variants:

* **`caption_general_a`** — the top 3–5 VoiceNet dimensions, **no emotion terms at all**.
  Emotion comes solely from the inline tags.
* **`caption_general_b`** — the same, plus the top 2–3 emotions.

**The >30 s split.** The scorer pads/truncates every input to exactly 30 s, so scoring a
45 s chain would silently describe only its first 30 s. Chains longer than 30 s are split
at **segment boundaries** into ≤30 s parts, each part scored, and the parts combined by
duration weight (VoiceNet regressions and emotion scores averaged; ordinal buckets taken
from the longest part, so bucket and label stay consistent). `n_score_parts` records how
many. In this tier **96.7 %** of chains are split — the normal case, not an edge
case.

**The emotion gate** is a tie-aware mid-rank ECDF, top 10 % *for that emotion*, max 3
named. This matters: the 40 emotion heads sit on different scales, and `emo_Awe` is at or
below zero for ~92 % of clips, so its p90 *value* is −0.0 and a naive `score >= p90_value`
test would name Awe on about a third of all clips. The mid-rank ECDF maps that tie block
to 0.47 and it correctly fails the gate. When nothing clears, **the clause is simply
absent** — there is no "no dominant emotion" string. `emotion_clause_present` records it.

Every ranking breaks ties **explicitly on the dimension or emotion name**, so captions are
byte-identical across runs.

## ⚠ Burst annotations mean two different things — read this

A `(parenthetical)` inside the transcript is **not** one claim. Measured across the corpus:

| source | paren & `n_bursts`==0 | paren & `n_bursts`>0 |
|---|--:|--:|
| **`vprof_vc`** | **37.7 %** | 11.4 % |
| `emolia` | 0.0 % | 28.5 % |
| `podcast` | 0.0 % | 44.0 % |
| `kartoffelphon` | 0.0 % | 24.5 % |
| `eurospeech` | 1.5 % | 46.2 % |
| `mls` | 0.0 % | 0.4 % |

In the **mined** corpora a parenthetical always coincides with a detected burst — it is
detector output, an observed event. In **`vprof_vc`** roughly two-fifths of clips carry a
parenthetical the detector never confirmed: those are burst *directions from the synthesis
prompt*, not observations.

Because `vprof_vc` is **100.0 %** of this tier, most parentheses here are the
unconfirmed kind. So each segment carries the two claims **separately**:

* `burst_detected` — `n_bursts > 0`, a real detection
* `burst_scripted` — a parenthetical with `n_bursts == 0`, requested but unconfirmed
* `burst_note` — `"detected"` / `"scripted-unconfirmed"` / `"none"`

**Do not treat them as the same signal.**

## ⚠ Language: the chain label is the FIRST clip only

`lang_iso` on a chain is taken from its first clip. That is **not** the language of the
whole chain. Measured on this tier by two independent methods (per-clip `lang` from the
index, and stopword detection on the segment text):

| | share of chains |
|---|--:|
| single-language | **7.4 %** |
| **mixes two languages** (de+en) | **92.6 %** |

The cause is structural, not a bug: a `vprof_vc` chain is five takes **by one cloned
voice of five different texts**, and those texts are not all in one language. The voice
identity is constant; the language is not.

So each chain carries the honest fields alongside the label:

* `langs` — sorted distinct languages actually present, e.g. `["de","en"]`
* `lang_mixed` — true when more than one
* `n_langs`
* per segment, `segments[i].lang`

**Filter on `langs` / `lang_mixed`, not on `lang_iso`**, unless you specifically want
"whatever the first clip was". The per-tier `en %` / `de %` tables above are computed on
the chain label and inherit exactly this caveat.

## Composition
| dataset | chains | licence |
|---|--:|---|
| `vprof_vc` | 319,650 | generated (see note) |
| `emolia` | 115 | CC-BY-4.0 |

This is the **CLEAN** release: `podcast` and `evasnippets` are removed for provenance, and `kartoffelphon` is held pending a licensing review.

## Files, audio and formats

**2,402 files at the repository root, in 800 numbered triples plus this card.**

| path | count | what it is |
|---|--:|---|
| `traj-t80-NNNNN.tar` | 800 | **the audio**, plus the per-chain JSON and the MOSS tokens |
| `traj-t80-NNNNN.parquet` | 800 | **the metadata**: one row per chain, 212 columns |
| `traj-t80-NNNNN.done` | 800 | per-shard verification counts written after the shard was sealed |

`NNNNN` runs `00000`..`00799`. Shard *n* of one kind always describes shard *n* of the other; the
last shard is short (165 chains, the rest hold 400) — 799 x 400 + 165 = **319,765**.

### Inside a tar — three members per chain, one stem

| member | format | detail |
|---|---|---|
| `<stem>.mp3` | **MPEG-1 Layer III, 48 kHz, 96 kbps CBR, mono** | the crossfaded chain, read out of the shipped bytes |
| `<stem>.json` | JSON | `chain_id`, both general captions, `script`, `segments`, `constant_descriptors`, `speaker_clause`, `tier_max`, `rules`, `dur_s`, `moss_frames`, `n_score_parts` |
| `<stem>.moss.npy` | `numpy` `uint16`, shape `[T, 12]` | MOSS-Audio-Tokenizer-v2 codes of the **rendered chain**, 12.5 fps, one frame = 12 tokens = 80 ms |

There is no `.vclap.npy` member here and no waveform format other than MP3.

### The join — parquet row to tar member

`shard` is the tar basename **without** the `.tar` extension, and the member stem is `chain_id`
with `|` replaced by `__`:

```python
tar_path    = row["shard"] + ".tar"              # "traj-t80-00000"  -> traj-t80-00000.tar
member_stem = row["chain_id"].replace("|", "__")  # "emolia|DE_x_W0,DE_x_W1,..." -> "emolia__DE_x_W0,DE_x_W1,..."
```

Verified on `traj-t80-00000`: the parquet's first three `chain_id` values map to the tar's first
three member groups, member for member.

Member names are long (a chain names all five of its source uids) and exceed the 100-byte POSIX
tar name field, so the tars use GNU long-name records. Every standard tar reader — Python's
`tarfile`, GNU `tar`, `webdataset` — handles this; a hand-rolled 512-byte header parser does not,
and will silently hand you truncated names.

## Columns

**212 columns per row.** The full generated reference, with a one-line meaning for every column, is
in **[`COLUMNS.md`](COLUMNS.md)**. The groups:

| group | columns | what it is |
|---|--:|---|
| identity and provenance | 9 | `chain_id`, `chain_key`, `shard`, `dataset`, `arm`, `uids`, `speaker`, `track`, `k` |
| language | 4 | `lang_iso` (**first clip only**), `langs`, `lang_mixed`, `n_langs` |
| audio and rendering | 5 | `dur_s`, `dur_s_soundfile`, `peak_clipped`, `lvl_spread_db`, `fades_ms` |
| MOSS tokens | 3 | `moss_frames`, `moss_n_vq`, `moss_frames_expected` |
| text | 3 | `script_text`, `segments` (the per-clip records), `constant_descriptors` |
| captions | 9 | `caption_general_a`/`_b` and the parts they were built from |
| emotion | 40 | `emo_*`, Empathic-Insight intensities **of the rendered chain** |
| voice character | 114 | `vn_<CODE>_reg` + `vn_<CODE>_bucket`, 57 VoiceNet dimensions |
| quality | 6 | 4 `qual_*` heads, `genuineness_0_6`, `blend_0_10` |
| vocal bursts | 2 | `n_burst_detected`, `n_burst_scripted` — **two different claims**, see below |
| trajectory qualification | 17 | `tier_max`, `qmax`, `cmax`, `C`, `rules`, `dim_a`/`dim_b`, the speaker cosines, the scoring-split fields |

**The 40 `emo_*`, 114 `vn_*` and 4 `qual_*` columns describe the RENDERED CHAIN**, scored on the
crossfaded audio, not averaged from the source clips. Per-clip values are inside `segments`.

## Loading

```python
import glob, json, io, tarfile
import numpy as np, soundfile as sf, pyarrow.parquet as pq

# metadata for the whole tier
meta = pq.read_table(sorted(glob.glob("traj-t80-*.parquet"))).to_pandas()
print(len(meta), "chains")

# the nested ladder is a filter, not a re-download
hard = meta[meta.tier_max >= 0.80]

# one chain, audio + codes + captions
with tarfile.open("traj-t80-00000.tar") as tf:
    names = [m.name for m in tf.getmembers() if m.name.endswith(".mp3")]
    stem  = names[0][:-4]
    wav, sr   = sf.read(io.BytesIO(tf.extractfile(stem + ".mp3").read()))
    codes     = np.load(io.BytesIO(tf.extractfile(stem + ".moss.npy").read()))
    info      = json.loads(tf.extractfile(stem + ".json").read())

print(codes.shape, sr)                      # (frames, 12) uint16, 48000
assert codes.shape[0] == int(len(wav)/sr*12.5)
print(info["caption_general_a"])            # voice only
print(info["caption_general_b"])            # voice + emotions
print(info["script"])                       # inline screenplay
for s in info["segments"]:
    print(s["start"], s["burst_note"], s["tag"], s["text"][:60])
```

## Selecting a tier

Every row of this release already satisfies `tier_max >= 0.80` — that is what the tier is. Within
it, `tier_max` (equivalently `qmax`, equal to it on every row here) is the selector:

```python
harder = meta[meta["tier_max"] >= 0.90]
```

**A correction to earlier prose.** Descriptions of this dataset have mentioned `tNN` and `pcNN`
selector columns. **They are not in this release.** Checked on the first, middle and last shard:
212 columns each, none of them `tNN` or `pcNN`. Those selectors belong to the upstream chain table
the tiers were packed from, where the distinction mattered — `tNN` nested by construction, `pcNN`
a per-cell packing that is *not* a subset of anything and must never be summed with or nested
inside `tNN`. Downstream of the packer only `tier_max` survives, and it is the nested one.

## Source datasets and licences

| source | what it is | licence |
|---|---|---|
| `emolia` | YODAS-derived emotional speech | CC-BY-4.0 |
| `kartoffelphon` | German speech (~51 % librivox) | CC-BY-4.0 · **held pending review** |
| `mls` | Multilingual LibriSpeech | CC-BY-4.0 |
| `eurospeech` | European parliamentary speech | CC-BY-4.0 |
| `vprof_vc` | synthetic voice profiles, Chatterbox-VC + SIDON | see note below |
| `snippets` | short reference snippets | CC-BY-4.0 |

`podcast` and `evasnippets` are **excluded from every public release** — their per-item
provenance is unknown or unclear and they are not openly redistributable.

`kartoffelphon` is excluded from the CLEAN releases **pending a licensing review, not for
any data defect**: ~51 % is librivox (public domain), the remainder is podcast-provenance
material. Its coverage is excellent (100 % `-id` speaker embeddings, the best of any
source). Every tier records `kartoffelphon_contribution` so the effect of clearing it is
visible without a rebuild.

Annotations are CC-BY-4.0. The generated voice-profile audio derives in part from EmoLia.

## Honest limitations

* **`vprof_vc` trajectories are constructed, not observed.** Takes are independent
  renditions of *different texts* by one cloned voice; the ordering is imposed by the
  packer. They carry no speaker-identity risk and no natural time axis. At the top of the
  ladder they are almost the entire corpus.
* **Five emotions cannot be trajectory endpoints under the two-sided rules** — Awe,
  Distress, Sadness, Disappointment, Helplessness. Their raw scores are so zero-inflated
  that the normalised value is effectively two-valued, so no per-step cap ≤0.25 can be
  satisfied. This is arithmetic, not missing data. `PXR` recovers four of the five.
* **EmoLia identity uses a timbre embedding**, not the verification model. Where
  `cos_source == "orange-tbr"` the 0.80 threshold is applied to
  `Orange/Speaker-wavLM-tbr`; store-to-store the equivalent of `-id >= 0.80` is
  `-tbr >= 0.788`, so this is fractionally *stricter*, but it is a different space and its
  pass rate reads ~7 points optimistic.
* **The emotion scorer is weak in absolute terms** — 8.0 % top-1 against the requested
  label on the voice profiles (chance 2.5 %). Selection is always on measured scores,
  never on a requested-emotion label.
* **Emotion clauses are common at chain level.** With 40 heads each gated at its own top
  decile, P(none clears) ≈ 0.9^40 ≈ 1.5 %, so ~98 % of chains name at least one emotion.
  That is the arithmetic of a 40-way gate, not evidence every chain is emotive.
* Language is `lang_iso_fixed`, post-correction. `other %` is dominated by the
  multilingual sources.

## Citation

```bibtex
@misc{laion_traj_t80,
  title  = {LAION Emotional-Trajectory Speech, tier T>=0.80},
  author = {LAION},
  year   = {2026},
  url    = {https://huggingface.co/datasets/laion/laion-emotional-trajectory-t80}
}
```