Not scraped from YouTube. Not recorded by anonymous crowds who don't speak the language. Every conversation in this dataset traces back to verified native speakers we know by name.

The [Human] Standard

Most speech datasets are built the same way: scrape the internet, hire anonymous contractors, run it through automated QC, ship it. The result? Models that are confidently wrong.

We built this dataset differently.

Native speakers only — Every recording made by verified speakers in their native language
Real conversations — Unscripted, natural dialogue between two people who actually know how to talk to each other
Channel-separated stereo — Each speaker on their own channel, ready for diarization training
Full traceability — Every file links back to the person who recorded it

This is the [human] standard for AI data.

Dataset Overview

Metric	Value
Total Duration	305 hours
Total Files	928
Total Size	~199 GB
Locales	18
Domains	Healthcare, Meetings, Call Centers
Format	WAV, PCM, Stereo
Sample Rate	44.1 kHz / 48 kHz
Speakers per File	2 (98.6% of files)
Transcripts	100% coverage, word-level timestamps

Audio Specifications

Format: WAV (PCM 16-bit / 24-bit lossless)
Sample Rate: 44.1 kHz or 48 kHz
Channels: Stereo (2-channel, speaker-separated)
Recording Platform: Zencastr / Riverside (professional remote recording)
SNR: 45-51 dB (studio-quality)
Speech Type: Natural, unscripted conversational dialogue
Typical Duration: 10-60 minutes per recording

All audio files are professionally recorded with channel separation — Speaker A on left channel, Speaker B on right channel. Ready for speaker diarization training out of the box.

Supported Languages

Locale	Language	Hours	Domain Focus
zh-CN	Chinese (Mandarin)	34h	Healthcare, Meetings
en-US	English (US)	31h	Healthcare, Meetings
pt-PT	Portuguese (Portugal)	30h	Healthcare
pt-BR	Portuguese (Brazil)	30h	Healthcare
es-MX	Spanish (Mexico)	28h	Healthcare, Meetings
en-GB	English (UK)	26h	Healthcare
fr-FR	French (France)	25h	Healthcare, Meetings
ja-JP	Japanese	20h	Healthcare, Meetings
ko-KR	Korean	18h	Healthcare
yue-HK	Cantonese	15h	Healthcare
de-DE	German	14h	Healthcare
it-IT	Italian	12h	Healthcare
es-ES	Spanish (Spain)	10h	Meetings
en-AU	English (Australia)	8h	Call Centers
en-IN	English (India)	6h	Call Centers
hi-IN	Hindi	5h	Call Centers
fr-CA	French (Canada)	4h	Call Centers
es-AR	Spanish (Argentina)	4h	Meetings

Domain Breakdown

Healthcare (205 hours)

Medical conversations across specialties: general practice consultations, mental health sessions, dental discussions, specialist referrals. Recorded with healthcare professionals and patients discussing real (simulated) medical scenarios.

Subdomains: General Medicine, Mental Health, Dentistry, Cardiology, Pediatrics, Geriatrics

Meetings (81 hours)

Business and professional conversations: team discussions, project planning, client calls, interview scenarios. Natural back-and-forth dialogue with interruptions, crosstalk, and real conversational dynamics.

Subdomains: Business Strategy, Project Management, Sales Calls, HR Interviews

Call Centers (19 hours)

Customer service scenarios: support calls, complaint handling, booking and scheduling. Simulated but realistic call center interactions with appropriate pacing and turn-taking.

Subdomains: Technical Support, Customer Service, Booking, Complaints

Transcription Details

Every audio file includes:

Full verbatim transcript — Including fillers, false starts, and disfluencies
Word-level timestamps — Start and end time for every word
Speaker diarization — Speaker labels (Speaker A / Speaker B) for every segment
Confidence scores — Per-word confidence from ASR model

Transcription Process:

Primary transcription via ElevenLabs Scribe v2
Secondary pass with Whisper Large v3
Human verification for samples

Format: JSON with structured segments

{
  "segments": [
    {
      "start": "00:00:01.240",
      "end": "00:00:05.890",
      "speaker": "Speaker A",
      "text": "Good morning, thanks for coming in today.",
      "words": [
        {"word": "Good", "start": 1.24, "end": 1.48, "confidence": 0.98},
        {"word": "morning", "start": 1.52, "end": 1.89, "confidence": 0.99}
      ]
    }
  ]
}

Dataset Creation Methodology

Recording Process

All recordings were conducted via professional remote recording platforms (Zencastr and Riverside) that capture each participant on separate audio tracks. This ensures:

Clean channel separation — No bleed between speakers
Consistent quality — Platform handles audio optimization
Natural conversation — Participants in comfortable environments

Contributor Selection

Contributors were recruited through our verified community network of 300,000+ members. Selection criteria:

Native speaker of target language
Clear speech without heavy regional accents (unless specifically required)
Comfortable with conversational topics
Passed audio quality screening

Quality Assurance

Multi-stage QC pipeline:

Automated checks: Duration, sample rate, channel count, SNR
Language verification: Confirmed correct language via speech recognition
Content review: Spot-checked for topic adherence and quality
Transcript validation: Compared ASR output against audio samples

Intended Uses

This dataset is designed for:

✅ Training and fine-tuning Automatic Speech Recognition (ASR) models
✅ Speaker diarization research and model development
✅ Conversational AI training (natural dialogue patterns)
✅ Healthcare NLP applications (medical conversation understanding)
✅ Multilingual ASR benchmarking
✅ Speech-to-text model evaluation
✅ Academic and commercial research

Out-of-Scope Uses

This dataset is not intended for:

❌ Real-time, safety-critical medical diagnosis systems
❌ Biometric speaker identification without consent
❌ Training voice cloning systems without proper licensing
❌ Any application that violates applicable privacy laws

Licensing

This dataset is available under a commercial license.

Pricing: $65 per hour of audio
Full dataset: ~$19,800
Custom subsets: Available (by locale, domain, or duration)
License type: Non-exclusive, perpetual use license

To discuss licensing, contact us directly or request access through this page.

Why UsergyAI?

We spent years inside the world's largest AI data companies. We saw how datasets actually get built — the shortcuts, the anonymous crowds, the "quality checks" that catch formatting errors but miss meaning.

Then we built something different.

300,000+ verified community members. Not anonymous contractors. Real people with real expertise who come back because we never lied to them about what a project paid or what it required.

Full traceability. Every data point traces back to a person we know. Not a user ID. A person.

The [human] standard. Because AI is only as good as its data, and data is only as good as the people who create it.

Contact

For licensing inquiries, custom datasets, or questions:

Swaroop (Founder)
📧 swaroop@usergy.ai
🌐 usergy.ai

The [human] standard for AI data.

Citation

@dataset{usergyai_conversational_speech_2026,
  author = {UsergyAI},
  title = {Global Conversational Speech Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/UsergyAI/Global-Conversational-Speech}
}

Downloads last month: 12

Total file size:

1.29 GB