SON

SON is a speech dataset of natural, unscripted conversations between friends. Each recording captures two speakers answering fun, party-game-style prompts together, producing spontaneous dialogue rich in interruptions, backchanneling, laughter, and other paralinguistic features. SON is particularly useful for building speech-to-speech (S2S) models for party games and social experiences, where models need to handle the energy, interruptions, and playful dynamics of real group interaction.

About the data

All conversations are recorded on participants' mobile devices. The audio is dual channel, with each speaker natively recorded on a separate channel. Audio files are delivered as WAV; samples in this dataset card are converted to MP3 for presentation.

If you're looking for studio-quality data, check out the emo-com dataset.

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.

Strengths

The primary strength of this dataset is its naturalness. Because these are real conversations between friends, the audio captures genuine full-duplex speech elements that are difficult to simulate or elicit in scripted settings. These include backchanneling, interruptions, overlapping speech, laughter, and other paralinguistic cues that reflect how people actually talk to each other.

Technical specs

Metric	Value
File format	WAV (MP3 in this dataset card)
Sample rate	48,000 Hz
Bit depth	16-bit
Segment duration	120 seconds each
Average SNR	21.74 dB
Average speech ratio	68.6%
Average spectral centroid	2.539 kHz
Average frequency content	11.581 kHz

Speaker metadata

Each recording includes demographic metadata for both speakers when available.

• Accent (inferred from IP address)
• Gender