README.md

---
pretty_name: MOSS · Emolia + Elise + Inline-Bursts (HQ, captioned)
license: other
language:
- de
- en
- multilingual
task_categories:
- text-to-speech
- audio-classification
- automatic-speech-recognition
size_categories:
- 1K<n<10K
tags:
- audio
- speech
- voice-acting
- expressive-tts
- emotion
- vocal-bursts
- moss-tts
- voicenet
- emonet
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.tar
- config_name: metadata
  data_files:
  - split: train
    path: metadata.parquet
- config_name: tokenized
  data_files:
  - split: train
    path: tokenized_audio.parquet
---

# MOSS · Emolia + Elise + Inline-Bursts — HQ, captioned

A **high-quality, richly captioned** slice of the MOSS-local voice-acting corpus: expressive speech
clips scored by a panel of acoustic detectors, filtered to the top by a composite reward, and captioned
in the **voice-acting** format (a "how the voice sounds / how to perform it" description plus the script
with inline vocal-burst tags). Audio is shipped both as **flac** (WebDataset tars) and as pre-computed
**MOSS-Audio-Tokenizer codes** for direct TTS training.

**6,263 clips** drawn from three sources:

| `source_dataset` | clips | what it is |
|---|---:|---|
| `emolia` | 3,768 | emotion-bucketed expressive speech (Emolia) |
| `mossinline` | 1,800 | speech with **inline vocal bursts** (laughs, sighs, gasps…) |
| `elise` | 695 | Elise / DramaBox-style dramatic delivery |

Every clip was scored on VoiceNet (57 perceptual voice dimensions), Empathic-Insight-Voice-Plus / EmoNet
(40 emotions + Arousal / Valence / Authenticity), a **genuineness** score (felt vs. performed), a
**vocal-burst blend / naturalness** score, and WER against the reference text; the composite `score`
(range 0.3–16.0, mean ≈ 10.4) selects the high-quality tail.

## Files

- **`data/data-000{0..3}.tar`** — WebDataset shards. Each member pair is `<sample_key>.flac` (audio) +
  `<sample_key>.json` (per-clip metadata). ~2,000 clips per shard.
- **`metadata.parquet`** — one row per clip with the full annotation set (scores + captions). Join to the
  audio via `sample_key`.
- **`tokenized_audio.parquet`** — the same clips as **MOSS-Audio-Tokenizer** codes, ready for MOSS-TTS
  training (no raw audio needed). Join via `key` (== `sample_key`).

## `metadata.parquet` columns

| column | meaning |
|---|---|
| `sample_key` / `id` | unique id, `"<source>__<local-id>"` (e.g. `emolia__…`); matches the tar member and the tokenized `key` |
| `source_dataset` | `emolia` / `mossinline` / `elise` |
| `text` | reference transcript |
| `ext` | audio extension (`flac`) |
| `vn_*` (57) | VoiceNet perceptual dimensions (warmth, roughness, tempo, register, resonance, …), ~0–6 scale |
| `ei_*` (43) | Empathic-Insight-Voice-Plus: 40 EmoNet emotions + `ei_Arousal`, `ei_Valence`, `ei_Authenticity` |
| `genu` | genuineness (felt vs. performed) |
| `blend` | vocal-burst blend / naturalness (0–10) |
| `bude_caption` | free-text BUD-E-Whisper caption |
| `inline_burst` | transcript with inline vocal-burst tags |
| `procedural_caption` | rule-based voice-acting caption: `GENERAL:` (how the voice sounds) + `SCRIPT:` (delivery cues + text) |
| `voice_acting_caption` | LLM-naturalised rewrite of the procedural caption (same structure, fluent wording) |
| `score` | composite selection reward (higher = higher quality) |

## `tokenized_audio.parquet` columns

| column | meaning |
|---|---|
| `key` | id (matches `sample_key`) |
| `target_codes` | MOSS-Audio-Tokenizer codes for the target clip, `int16` bytes, shape `[target_frames, n_codebooks]` |
| `target_frames` | number of code frames |
| `ref_codes` / `ref_frames` | optional reference-voice codes (often empty) |
| `text` | reference transcript |
| `procedural_caption`, `voice_acting_caption` | as above |
| `source_dataset` | `emolia` / `mossinline` / `elise` |

## Usage

Stream the audio + captions (WebDataset):

```python
from datasets import load_dataset
ds = load_dataset("TTS-AGI/moss-emolia-elise-hq-captioned", split="train", streaming=True)
ex = next(iter(ds))
print(ex["json"]["voice_acting_caption"])
ex["flac"]["array"], ex["flac"]["sampling_rate"]
```

Load just the scores + captions:

```python
ds = load_dataset("TTS-AGI/moss-emolia-elise-hq-captioned", "metadata", split="train")
```

Decode the MOSS tokens for training:

```python
import numpy as np, pandas as pd
df = pd.read_parquet("tokenized_audio.parquet")
row = df.iloc[0]
codes = np.frombuffer(row["target_codes"], np.int16).reshape(row["target_frames"], -1)
```

## Notes

- Captions and scores are **model-generated** (VoiceNet / EmoNet / genuineness / blend detectors +
  procedural templating + LLM rewrite) and are not manually verified.
- Part of the **MOSS-local voice-acting** data-generation effort. `license: other` — see the source
  datasets for provenance and terms.