CoolFace
Datasetpublic

susameddin/Sympatheia-18k

Sympatheia-18k Sympatheia-18k is an emotion-aware spoken dialogue dataset for empathetic speech synthesis research. It contains 18,000 query–response pairs across 12 emotion categories, each accompanied by synthesized audio and text transcripts. Dataset Structure Subset Unique Queries Responses Description Emotional 8,400 train / 3,600 eval 8,400 train / 3,600 eval Emotional queries with emotionally-matched responses Neutral 350 train / 150 eval 4,200… See the full description on the dataset page: https://huggingface.co/datasets/susameddin/Sympatheia-18k.

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face
2likes203downloads
Dataset Card

Sympatheia-18k

Sympatheia-18k is an emotion-aware spoken dialogue dataset for empathetic speech synthesis research. It contains 18,000 query–response pairs across 12 emotion categories, each accompanied by synthesized audio and text transcripts.

Dataset Structure

SubsetUnique QueriesResponsesDescription
Emotional8,400 train / 3,600 eval8,400 train / 3,600 evalEmotional queries with emotionally-matched responses
Neutral350 train / 150 eval4,200 train / 1,800 evalNeutral queries each paired with 12 emotion-targeted responses

Files

  • —{Emotional,Neutral}/audio/{train,eval}/ : WAV audio files named {LABEL}_{INDEX}_{EMOTION}.wav
  • —{Emotional,Neutral}/metadata/text_pairs_{train,eval}.jsonl : query–response text pairs with query_style and user_emotion labels (see column descriptions below)
  • —encoded_{train,eval}.jsonl : audio tokens pre-encoded using the GLM-4-Voice decoder for use with the Sympatheia model, along with emotion labels and assigned valence–arousal values per emotion class
  • —{Emotional,Neutral}/metadata/emotion_va_mapping.json : per-class emotion → valence–arousal mapping (programmatically defined, not human-annotated)

Column Descriptions

Each text_pairs entry contains two emotion-related columns:

  • —`query_style` : the TTS rendering style used to synthesize the query audio. In the Emotional subset this reflects the speaker's emotional delivery; in the Neutral subset this is always Neutral.
  • —`user_emotion` : the underlying emotion attributed to the user, which determines the tone and content of the response. In the Emotional subset this is identical to query_style. In the Neutral subset this is the inferred user emotion despite the neutral delivery (e.g. a neutrally-spoken message from someone who is actually Frustrated), and is the primary signal used to generate an appropriate empathetic response.

Related Project

This dataset was created for Sympatheia, an emotion-aware empathetic speech dialogue system.

Emotion Categories

12 discrete emotion classes: Angry, Anxious, Content, Disgusted, Excited, Frustrated, Happy, Neutral, Relaxed, Sad, Surprised, Tired