| # MultiEmo-Test |
|
|
| MultiEmo-Test is an English evaluation set for instruction-following multi-emotion text-to-speech synthesis. It accompanies [HybridEmo](https://github.com/ictnlp/HybridEmo), a system for modeling sequential emotion trajectories and simultaneous emotion blending within an utterance. |
|
|
| The dataset is intended for evaluation only. It contains synthesis text, natural-language emotion instructions, emotion annotations, and prompt audio for speaker-timbre conditioning. It does not contain target synthesized speech. |
|
|
| ## Dataset Composition |
|
|
| MultiEmo-Test contains 720 examples in a single test split. |
|
|
| | Task | Subset | Number of examples | Description | |
| |---|---:|---:|---| |
| | Emotion trajectory | 1E | 200 | A single emotion is expressed throughout the utterance. | |
| | Emotion trajectory | 2E | 200 | Two emotions are expressed sequentially. | |
| | Emotion trajectory | 3E | 200 | Three emotions are expressed sequentially. | |
| | Emotion blending | 2E blending | 120 | Two emotions are expressed simultaneously. | |
| | **Total** | | **720** | | |
|
|
| The trajectory subset uses seven emotion labels: |
|
|
| `angry`, `disgusted`, `fearful`, `happy`, `neutral`, `sad`, and `surprised`. |
|
|
| The blending subset uses six emotion labels: |
|
|
| `angry`, `disgusted`, `fearful`, `happy`, `sad`, and `surprised`. |
|
|
| ## Data Format |
|
|
| The dataset is distributed as a JSON Lines manifest and a directory of prompt audio files: |
|
|
| ```text |
| MultiEmo-Test/ |
| ├── README.md |
| ├── test.jsonl |
| └── prompt-audio/ |
| └── *.wav |
| ``` |
|
|
| Each line in `test.jsonl` contains the following fields: |
|
|
| | Field | Type | Description | |
| |---|---|---| |
| | `id` | integer | Unique example identifier. | |
| | `trajectory_type` | integer or null | Number of stages for a trajectory example (`1`, `2`, or `3`); null for blending examples. | |
| | `emotion_trajectory` | list of strings or null | Ordered emotion labels for a trajectory example. | |
| | `blended_type` | string or null | Set to `blended` for blending examples; null for trajectory examples. | |
| | `blended_emotion` | list of strings or null | Two emotion labels to be expressed simultaneously. | |
| | `text` | string | Text to synthesize. | |
| | `instruction` | string | Natural-language instruction describing the intended emotional expression. | |
| | `prompt_audio` | string | Relative path to the speaker-timbre reference audio. | |
| | `prompt_text` | string | Transcript of the prompt audio. | |
| | `prompt_key` | string | Source key associated with the prompt audio. | |
|
|
| Example trajectory record: |
|
|
| ```json |
| { |
| "id": 1, |
| "trajectory_type": 1, |
| "emotion_trajectory": ["angry"], |
| "blended_type": null, |
| "blended_emotion": null, |
| "text": "Can you believe the audacity of that person?", |
| "instruction": "Read this passage with a consistently angry tone.", |
| "prompt_audio": "prompt-audio/common_voice_en_509177.wav", |
| "prompt_text": "The autonomous ship floated closer to receiving its flying cargo.", |
| "prompt_key": "yuekai/seed_tts_cosy2::path::common_voice_en_509177.wav" |
| } |
| ``` |
|
|
| ## Loading the Dataset |
|
|
| After downloading the repository, the manifest can be loaded with the Hugging Face `datasets` library: |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset( |
| "json", |
| data_files={"test": "test.jsonl"}, |
| )["test"] |
| |
| example = dataset[0] |
| print(example["instruction"]) |
| print(example["prompt_audio"]) |
| ``` |
|
|
| The value of `prompt_audio` is relative to the dataset root. For example, `prompt-audio/example.wav` should be resolved from the directory containing `test.jsonl`. |
|
|
| ## Intended Use |
|
|
| MultiEmo-Test is designed to evaluate whether instruction-following TTS systems can: |
|
|
| - maintain a specified emotion throughout an utterance; |
| - follow two- or three-stage emotion trajectories in the requested order; |
| - express two target emotions simultaneously; |
| - preserve the speaker timbre provided by the prompt audio. |
|
|
| The dataset is not intended as a training corpus or as a comprehensive representation of all emotions, emotion transitions, languages, speakers, or real-world speaking conditions. |
|
|
| ## License |
|
|
| MultiEmo-Test is released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite the HybridEmo paper: |
|
|
| ```bibtex |
| @misc{zhou2026sequentialtrajectoriessimultaneousblending, |
| title={Sequential Trajectories and Simultaneous Blending: Multi-Emotion Modeling for Instruction-Following TTS}, |
| author={Yan Zhou and Yun Hong and Yang Feng}, |
| year={2026}, |
| eprint={2608.30325}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CL}, |
| url={https://arxiv.org/abs/2608.30325}, |
| } |
| ``` |
|
|
| Project repository: [https://github.com/ictnlp/HybridEmo](https://github.com/ictnlp/HybridEmo) |
|
|