Emo-con is a speech dataset of real emotional conversations between people actively supporting each other. Unlike most emotion datasets, which rely on acted, pseudo-acted, or scripted speech, Emo-con captures genuine emotional expression in the context of mutual support. It reflects the way people actually talk to a close friend when sharing hardships, daily experiences, and jokes.

We operate support and community groups with licensed professionals, so we can assume all participants are attending to get some kind of emotional support or bonding experience. Most participants already know each other, but even the strangers get along well the first time they meet.

Conversations range from grief and feelings of being trapped or exhausted to laughing about silly things. We collect data "in the wild," minimizing the unnatural awkwardness that often results when contributors are set up to speak with each other in job-like settings. The people in conversation are navigating real moments of emotional connection, offering comfort and sharing feelings and experiences. The speaker pool includes women and men ages 18 to 84.

Current voice models lack emotional intelligence, as shown by Good-Pal-Bench. Emo-con aims to teach models to be emotionally supportive and to converse the way a close friend would, responding with appropriate tone, timing, and content. The dataset includes a lot of backchannels, interruptions, and active turn-taking, making it particularly well-suited for training full-duplex S2S models.

This is an ongoing data collection effort, and the samples here represent an early release as the dataset continues to grow. The data may be used for commercial purposes.

Transcripts

Each recording includes human-verified, full verbatim transcripts with speaker diarization and audio tags (e.g. [laughing]) to capture paralinguistic events. Transcript samples are available in the dataset viewer.

Transcription Process

All transcripts go through a minimum of two people. A transcriber works through the audio from scratch using a custom-built transcript editor, producing a full verbatim transcript with precise timestamps down to the millisecond, labeled speaker turns for diarization, and audio event tags for non-speech sounds (e.g. [phone buzzing], [laughing], [door closing]).

Once the initial transcription is complete, the transcript is run through an automated format and spelling checker. The transcriber reviews and fixes any detected errors before submitting. This version is then passed to a senior reviewer, someone with a proven track record of high-quality transcripts, who listens to the audio in its entirety and manually corrects any remaining spelling errors or inconsistencies by hand.

All transcription is handled in-house.

Recording method

Every speaker records on a Yeti Nano studio microphone with a pop filter (although this can be changed upon request). The source audio is natively recorded with separate speaker channels, preserving each voice in isolation. The samples presented here are stereo conversations with each speaker on a separate channel, time-aligned to preserve the natural conversational dynamics.

Speaker metadata

Each speaker has the following metadata collected alongside their recordings.

Gender
Age range
City where they grew up
Race
Ethnicity

Audio quality and technical analysis

All recordings are studio quality. Averages across the samples in this release.

Sample rate. 48,000 Hz
Bit depth. 24-bit
File format. WAV (converted to MP3 for presentation, but will be provided as WAV)
Mean SNR. 16.92 dB
Median RMS. 0.05951
Average speech ratio. 0.792
Spectral centroid. 3.397 kHz
Frequency content. 19.921 kHz

Downloads last month: 90

Total file size:

469 MB