README.md

---
language:
  - en
  - zh
license: cc-by-nc-4.0
pretty_name: MoVE — Multilingual Voice Emotion (paired EN/ZH)
size_categories:
  - 100K<n<1M
task_categories:
  - audio-to-audio
  - text-to-speech
  - translation
tags:
  - speech-translation
  - emotional-tts
  - paired-speech
configs:
  - config_name: "1hr"
    data_files:
      - split: train
        path: metadata_1hr.tsv
  - config_name: "50hr"
    data_files:
      - split: train
        path: metadata_50hr.tsv
  - config_name: "100hr"
    data_files:
      - split: train
        path: metadata_100hr.tsv
  - config_name: "500hr"
    data_files:
      - split: train
        path: metadata_500hr.tsv
  - config_name: "1000hr"
    data_files:
      - split: train
        path: metadata_1000hr.tsv
---

# MoVE — Multilingual Voice Emotion (paired EN/ZH)

Repo: [`47z/MoVE`](https://huggingface.co/datasets/47z/MoVE)

Paired English / Chinese emotional speech for speech-to-speech translation
training. Every row is a `(en_wav, zh_wav)` pair belonging to one of five
emotion categories: **angry, crying, happy, laugh, sad**.

## Subsets (nested, emotion-balanced)

| subset | en hours | pairs   | per-emotion hours |
|-------:|---------:|--------:|------------------:|
| 1hr    | 1.00     | 839     | 0.20 each         |
| 50hr   | 50.00    | 42,467  | 10.00 each        |
| 100hr  | 99.99    | 85,004  | 20.00 each        |
| 500hr  | 500.00   | 425,952 | 100.00 each       |
| 1000hr | 1055.67  | 896,477 | 95–133 (full)     |

Smaller subsets are strict subsets of larger ones. Hours are measured by
**English** audio duration; the Chinese side is paired one-to-one and
~5% longer in total.

## Repository layout

```
47z/MoVE/
├── README.md
├── metadata_1hr.tsv       # 5 columns: zh_path, en_path, zh_text, en_text, category
├── metadata_50hr.tsv
├── metadata_100hr.tsv
├── metadata_500hr.tsv
├── metadata_1000hr.tsv    # full
├── metadata_all.tsv       # all 896k pairs + subset column
├── data/                  # audio packed into tar shards
│   ├── s1hr_000.tar
│   ├── s50hr_*.tar
│   ├── s100hr_*.tar
│   ├── s500hr_*.tar
│   └── s1000hr_*.tar
└── scripts/
    └── tsv_to_metadata.py   # TSV -> Kimi-Audio training JSONL
```

Inside each tar, paths preserve the original tree:

```
en/<emotion>/<id>_en.wav
zh/<emotion>/<id>_zh.wav
```

So `tar -xf data/*.tar` reproduces a flat `en/` + `zh/` directory layout.

## Quick start

### Download just one subset (e.g. 100hr ≈ 34 GB audio)

```bash
huggingface-cli download 47z/MoVE \
  --repo-type dataset --local-dir ./move \
  --include "metadata_100hr.tsv" \
            "data/s1hr_*.tar" \
            "data/s50hr_*.tar" \
            "data/s100hr_*.tar" \
            "scripts/*"

cd move
for f in data/*.tar; do tar -xf "$f"; done   # extracts en/ and zh/
```

The wav paths in `metadata_100hr.tsv` (e.g. `en/happy/happy_000001_en.wav`)
now resolve to local files.

### Download the full 1000hr (~340 GB)

```bash
huggingface-cli download 47z/MoVE \
  --repo-type dataset --local-dir ./move
cd move && for f in data/*.tar; do tar -xf "$f"; done
```

### Required tar shards per subset

| subset  | tar prefixes to download |
|---------|--------------------------|
| 1hr     | `s1hr_*` |
| 50hr    | `s1hr_*`, `s50hr_*` |
| 100hr   | `s1hr_*`, `s50hr_*`, `s100hr_*` |
| 500hr   | `s1hr_*`, `s50hr_*`, `s100hr_*`, `s500hr_*` |
| 1000hr  | all (`s*_*`) |

## metadata.tsv schema

Each `metadata_<size>.tsv` is tab-separated with this header:

| column   | example | description |
|----------|---------|-------------|
| zh_path  | `zh/happy/happy_000001_zh.wav` | path to Chinese wav (relative) |
| en_path  | `en/happy/happy_000001_en.wav` | path to English wav (relative) |
| zh_text  | `父亲。一个人出去约会...`         | Chinese transcript |
| en_text  | `Dater. A person who...`        | English transcript |
| category | `happy`                         | one of {angry, crying, happy, laugh, sad} |

### metadata_all.tsv (overview)

Same 5 columns plus one `subset` column:

| column   | values | meaning |
|----------|:------:|---------|
| `subset` | `1hr` / `50hr` / `100hr` / `500hr` / `1000hr` | the **smallest** subset this row belongs to |

Subsets are nested, so a row with `subset="1hr"` is also in 50hr, 100hr,
500hr and 1000hr; `subset="1000hr"` means the row is only in the full
dataset. Subset sizes (cumulative):

| subset value | cumulative includes | total rows |
|--------------|---------------------|-----------:|
| `1hr`        | self                | 839 |
| `50hr`       | 1hr + 50hr          | 42,467 |
| `100hr`      | 1hr + 50hr + 100hr  | 85,004 |
| `500hr`      | + 500hr             | 425,952 |
| `1000hr`     | full                | 896,477 |

Filter with pandas:

```python
import pandas as pd
df = pd.read_csv("metadata_all.tsv", sep="\t")

# Build any subset by including this tier and all smaller ones:
order = ["1hr", "50hr", "100hr", "500hr", "1000hr"]
def subset(df, target):
    keep = order[:order.index(target) + 1]
    return df[df["subset"].isin(keep)]

df_50hr = subset(df, "50hr")   # equivalent to metadata_50hr.tsv
```

## Convert to Kimi-Audio training format

The included script `tsv_to_metadata.py` converts any metadata TSV
into the conversation-format JSONL used to replicate the
[Kimi-Audio](https://github.com/MoonshotAI/Kimi-Audio) speech-to-speech
training pipeline described in our paper.

```bash
python scripts/tsv_to_metadata.py metadata_100hr.tsv \
  -o metadata_shard.jsonl
```

Output schema (one JSON object per line):

```json
{
  "task_type": "s-s",
  "conversation": [
    {"role": "user", "message_type": "text", "content": "Translate the given English speech into Chinese while preserving its expressiveness."},
    {"role": "user", "message_type": "audio", "content": "en/happy/happy_000001_en.wav"},
    {"role": "assistant", "message_type": "audio-text", "content": ["zh/happy/happy_000001_zh.wav", "中文文本..."]}
  ]
}
```

Each TSV row produces 2 conversations by default (en→zh + zh→en). Use
`--directions en2zh` / `--directions zh2en` to keep one direction only.

## How it was built

- En/Zh paired by the `<id>_en` / `<id>_zh` suffix on the original
  `entries_shard_*.jsonl` records (2,801 unpaired entries dropped).
- Tier assignment per emotion: shuffle (seed=42), accumulate English
  duration, label each pair with the smallest subset it falls into.
- Tar shards are emotion-balanced (round-robin) within each tier and
  capped to keep individual tars manageable (≤ ~20 GB).

## Citation

```bibtex
@article{chen2026move,
  title={MoVE: Translating Laughter and Tears via Mixture of Vocalization Experts in Speech-to-Speech Translation},
  author={Chen, Szu-Chi and Tsai, I-Ning and Lin, Yi-Cheng and Huang, Sung-Feng and Lee, Hung-yi},
  journal={arXiv preprint arXiv:2604.17435},
  year={2026}
}
```

## License

CC BY-NC 4.0