Video Interview Speech Dataset

26,000+ hours of consented, unscripted video interview recordings. Single-speaker spontaneous speech with ASR transcripts, timestamped at the sentence level. 205 accents across 31 languages.

26,000+
Licensable Hours
97,000+
Unique Speakers
205
Accents
31
Languages
107K+
Interviews

Sample Preview

60-second excerpt from a US English interview. Speaker discusses a workplace scenario โ€” typical of the dataset's spontaneous, unscripted speech style.

๐ŸŽฅ Video

๐Ÿ”Š Audio (WAV)

16 kHz ยท mono ยท PCM 16-bit

๐Ÿ“ Transcript Excerpt (with timestamps)

[00:00.620 โ†’ 00:06.100] Yes, I've worked in different positions where I had to promote a tool. [00:06.580 โ†’ 00:14.320] I'll go back to my time when I was working for [REDACTED] and I was creating the PMO from the ground up. [00:14.520 โ†’ 00:26.000] And I had to implement a brand new tool called AHA that would act as a single source of truth and a system of record for project managers. [00:26.640 โ†’ 00:27.980] They didn't like it. [00:28.720 โ†’ 00:40.380] They were used to doing things their own way. But being the forefront of this project, I had to educate them of the benefit they will derive when utilizing this tool, how to save time, [00:40.540 โ†’ 00:48.120] how to streamline progress reports for them, how it will help them know who in their team are [00:48.120 โ†’ 00:56.760] underutilized when they're doing utilization reports, and also how this particular tool will [00:56.760 โ†’ 01:00.000] just make their life so easy.

Dataset Specifications

Content Type
Job interview recordings โ€” one person answering questions
Speech Style
Unscripted, spontaneous answers (not read from a script)
Speakers per Recording
1 (interviewer is AI-generated and not recorded)
Video Resolution
640ร—480 (SD webcam capture)
Audio Sample Rate
48 kHz (stereo or mono, single voice)
Audio SNR
Median 37 dB (83% of clips โ‰ฅ 25 dB)
Video Codec
H.264 (MP4), some VP8/VP9 (WebM)
Frame Rate
~15 fps median (variable, browser capture)
Voice Activity
93% median (speech share of recorded time)
Mouth Visible
95% of seconds (both mouth corners detected)
Median Clip Length
54 seconds (p90: 122 s, max: 30 min)
Full Catalogue Size
~7.6 TB (301 MB/hour with video + audio)

Transcripts & Annotations

Transcript Coverage
99.7% of hours carry an ASR transcript
Timestamps
Sentence-level start/end times (98.7% coverage)
Transcript Type
Machine-generated (ASR) โ€” not human-verified
Language Label Accuracy
95% (validated against Whisper on 230 random clips)

Language & Accent Coverage

91.6% English with global accent diversity. 30+ non-English languages with 1+ hour each. 20 accents exceed 100 hours.

Top Represented Regions

RegionKey CountriesLanguages
North AmericaUnited States, CanadaEnglish
South AsiaIndia, Pakistan, BangladeshEnglish, Hindi, Urdu, Bengali
AfricaNigeria, Kenya, Ghana, South Africa, EgyptEnglish, French, Arabic
EuropeUnited Kingdom, Germany, France, PolandEnglish, German, French, Polish
Middle EastUAE, IsraelEnglish, Arabic, Hebrew
Southeast AsiaPhilippines, SingaporeEnglish, Tagalog
South AmericaArgentina, BrazilSpanish, Portuguese

Use Cases

๐Ÿ—ฃ๏ธ

ASR & Speech Recognition

Train and evaluate speech-to-text models on diverse accented English and 30+ other languages.

๐ŸŽฏ

Speaker Diversity & Diarization

97,000+ unique speakers for training diarization, speaker segmentation, and accent-classification models.

๐Ÿ‘„

Lip Reading & Visual Speech

95% mouth visibility. SD video with aligned audio and transcripts for audio-visual speech models.

๐ŸŒ

Accent & Dialect Research

205 accent categories (language ร— country). Study accent variation at scale across global populations.

๐Ÿ“Š

NLP & Conversation Analysis

Unscripted professional speech for language modeling, fluency analysis, and workplace NLP.

๐Ÿ”Š

Audio Quality & Enhancement

Real-world SNR distribution (median 37 dB) for noise robustness benchmarking and enhancement models.

Catalogue Growth

Coverage Period
June 2025 โ€” present (continuously growing)
Current Run Rate
~2,300 new hours per month
Licensing Model
All hours are consented and ready to license โ€” no collection delay

Consent & Licensing

Speaker Consent
Every recording from a speaker who consented to data sharing
Privacy
No speaker names, emails, or employer names in delivered data
Restrictions
Voiceprints, faceprints, facial animation, lip-sync avatars, and synthetic voice cloning are outside the consent scope
Opt-Out
Speakers who opted out after consent are excluded from the catalogue
โฌ‡ Download Sample Clips ๐Ÿ“ง Contact Us About This Dataset โ† Back to All Samples

Need a Custom Subset?

Filter by language, accent, region, clip length, or audio quality. We can deliver exactly the slice you need.

Request Custom Subset