DualEmotionBench

---

1. Dataset Summary (English)

The DualEmotionBench dataset is specifically designed to evaluate TTS models' performance in intra-utterance emotion-transition synthesis and bilingual speech generation.

In this benchmark, an emotion transition denotes a directional temporal evolution from a starting emotion to an ending emotion within a single utterance (e.g., happiness -> sadness), rather than a blended or co-occurring emotional state.

Each audio sample is carefully curated to feature emotional transitions within a single sentence. In addition to full-sentence audio, we provide segmented annotations based on emotional transition points.

2. Data Fields

The dataset is indexed via .jsonl files and includes the following key fields:

• id: Unique utterance identifier.
• original_emotion_pair: The target emotion pair contained within the sentence (e.g., happiness, sadness).
• text: The transcript of the speech.
• audio: Path to the full-sentence audio (supports online preview on Hugging Face).
• split_results: A list of segmented fragments, each containing specific text and its corresponding emotion. Fragments are stored in temporal order; their emotion sequence defines the directional transition.
• split_audio_prefix: The path prefix for the segmented audio clips.

3. Emotion Tags

The dataset covers 7 core emotion labels:

• neutral
• happiness
• sadness
• anger
• fear
• disgust
• surprise

4. Usage

You can load the dataset via the datasets library in Python:

from datasets import load_dataset

# Load Chinese subset
dataset_cn = load_dataset(
    "kunge123/DualEmotionBench",
    data_files="metadata_cn.jsonl"
)

# Load English subset
dataset_en = load_dataset(
    "kunge123/DualEmotionBench",
    data_files="metadata_en.jsonl"
)