North America · United States
English (US) conversational speech dataset
1,500 hours of spontaneous English (US) conversations recorded in the United States across 150 vetted native speakers. Dual-channel audio with time-aligned transcripts, speaker metadata (age, gender, region, device), and full chain-of-consent documentation. Licensed for commercial ASR, TTS, and voice AI training.
Dataset specifications
| Language | English (US) |
|---|---|
| Recorded in | United States |
| Speech type | Spontaneous two-person conversations |
| Volume | 1,500 hours |
| Speakers | 150 vetted native speakers |
| Audio | Dual-channel (one channel per speaker) |
| Transcripts | Time-aligned, human-verified |
| Metadata | Age, gender, region, recording device |
| License | Commercial (training & evaluation) |
| Price | $90 per hour of audio |
Need more hours, another dialect or your own recording setup? We collect speech data to spec.
Built for
Training and evaluating speech models that must handle real English (US) speech:
- Automatic speech recognition (ASR / STT)
- Text-to-speech & voice synthesis prosody
- Voice agents and IVR for the United States
- Speaker diarization & turn-taking models
- Speech LLM / audio-understanding fine-tuning
Why conversational data?
Models trained on read speech consistently underperform on spontaneous audio. Disfluencies, overlap, and natural speech rates aren’t in the training distribution. This dataset closes that gap for English (US). Read more in our guide to conversational vs. scripted speech data.
Free samples
Audit the English (US) data yourself
Share your company email and we’ll email you representative audio, matching transcripts, and the metadata schema. Check audio quality, transcript accuracy, and speaker diversity before you spend a dollar.
Public alternatives
Public US English speech data vs. this dataset
| Source | What it is | License | Commercial training |
|---|---|---|---|
| Switchboard-1 Release 2 (LDC97S62) (opens in a new tab)Linguistic Data Consortium (LDC) | Two-sided US English telephone conversations collected by Texas Instruments in 1990-1991. | LDC User Agreement for Non-Members (fee applies; noncommercial research and education only); commercial use requires paid LDC for-profit membership | Paid license |
| Fisher English Training Speech Part 1 (LDC2004S13) (opens in a new tab)Linguistic Data Consortium (LDC) | English conversational telephone speech created at LDC during 2003, two-channel audio. | LDC User Agreement for Non-Members (fee applies; noncommercial research and education only); commercial use requires paid LDC for-profit membership | Paid license |
| The People's Speech (opens in a new tab)MLCommons | Large transcribed English speech set built from audio downloaded via the archive.org API. | CC BY and CC BY-SA subsets (versions 2.0 to 4.0) | Allowed |
| Mozilla Common Voice English (Scripted and Spontaneous Speech) (opens in a new tab)Mozilla Common Voice (via Mozilla Data Collective) | Volunteers read English sentences aloud, with curated American English subsets; a separate Spontaneous Speech release has unscripted answers to questions. | CC0 1.0 | Allowed |
Switchboard and Fisher are telephone calls from 1990-1991 and 2003 under LDC licensing; this dataset is US two-person conversation with a commercial license and explicit consent from every speaker.
Licenses as published by each source, checked October 2026. Check the current terms before use.
FAQ
Frequently asked questions
What does the English (US) speech dataset include?
1,500 hours of spontaneous two-person conversations from 150 vetted native speakers recorded in the United States, delivered as dual-channel audio with time-aligned transcripts, speaker metadata (age, gender, region, device), and consent documentation.
What makes US English hard for speech recognition?
US English varies more than most models expect. Southern, New York, Midwestern and African American English differ in vowels, rhythm and grammar, and casual speech is full of reductions like "gonna" and "wanna". Models trained mostly on read or broadcast audio lose accuracy on exactly these features in real conversations.
Is there free US English speech data?
Yes, the People's Speech and Mozilla Common Voice English are free for commercial use. Switchboard-1 and Fisher English Training Speech Part 1 need a paid license. This dataset is spontaneous two-person conversation recorded in the United States, with speaker metadata, one channel per speaker and signed commercial-use consent.
Can I use this English (US) dataset commercially?
Yes. The dataset is licensed for commercial use, including training ASR, TTS, and other voice-AI models. Every speaker has signed explicit commercial-use consent, and we provide the full consent chain for your legal review.
How much does the English (US) speech dataset cost?
Pricing is $90 per hour of audio. Volume discounts are available for full-dataset and multi-language purchases. Contact us for a quote.
Can I evaluate samples before buying?
Yes. Request samples with your company email, and we'll email you representative audio files and matching transcripts so your team can audit quality before committing.
Related datasets
You might also need
$60/hr