Note: The audio samples presented here have been compressed to MP3 for browser playback and do not reflect the actual acoustic quality of the dataset. Refer to the Technical Specs below for the specifications of what will actually be delivered.

Education Subject Samples is a speech dataset of real tutoring conversations between tutors and people seeking help on various academic subjects. The dataset is captured in real-life settings, covering a variety of subjects including math, chemistry, computer science, and english prep for standardized tests. The speaker pool includes both male and female speakers.

As speech-to-speech models become more capable, one of the most natural use cases is education. Many people will converse with AI like it is a tutor. This dataset can be useful for expanding adoption of S2S in specific domains such as education, where natural conversational dynamics, subject-specific terminology, and the back-and-forth of real instruction are essential for building effective models.

The conversations include a lot of interruptions, backchanneling, and paralinguistic features, reflecting the natural dynamics of real tutoring sessions where speakers ask clarifying questions, and react in real time.

Transcripts

This dataset includes human-verified transcripts with speaker diarization. Transcripts are full verbatim, with audio tags (e.g. [laughing]) to capture paralinguistic events.

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

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 conversation includes per-speaker metadata for both participants.

Gender
Age range
City where they grew up (accent info)
Race
Ethnicity

Each conversation also has a single subject label for the topic discussed.

Technical Specs

Metric	Value
File format	Delivered as WAV (converted to MP3 for presentation here)
Average sample rate	48,000 Hz
Average bit depth	24
Average SNR	17.96 dB
Average RMS	0.018702
Average speech ratio	0.415
Average spectral centroid	2.652 kHz
Average frequency content	18.74 kHz

Downloads last month: 27

Total file size:

11.7 MB