Datasets:
US English Speech Dataset
965 hours of unscripted conversation between American English speakers, recorded between 2 March and 3 April 2026: 1,517 remote conversations among 180 speakers, each speaker on their own device and channel, 38 minutes long on average. Every conversation has a machine transcript with speaker labels and word-level timestamps.
In stock. Download access within 24 hours of a signed order. The figures on this card were counted in our recording database on 18 November 2026.
This repository is the dataset card only. The audio is licensed commercially and isn't hosted on Hugging Face. Excerpts with their uncorrected machine transcripts play on the English (US) dataset page, where you can also ask for the full sample recordings. To evaluate a purchase, request a written quote so the team can confirm available recordings, permitted use and total price for your brief.
Specifications
| Language | US English |
| Speech | 965 hours of two-person conversation |
| Volume | 965 hours in 1,517 recordings |
| Speakers | 180 speakers, each of whom gives American English as their dialect; every conversation is between two of them. |
| Hours by country | The speakers were chosen by dialect: all 180 give American English as theirs. Gender and age are listed below for the speakers who stated them. |
| Recorded | 2 March 2026 to 3 April 2026 |
| Audio | WAV at 48 kHz and 24 bits, one microphone per speaker. Each conversation comes as a stereo file with one speaker per channel plus a separate file per speaker. |
| Devices | About 90% recorded on a laptop and 9% on a phone, each speaker on their own device in a quiet place. |
| Transcripts | Machine transcripts for every conversation, not reviewed by a person: speaker labels, segment start and end times and word-level timestamps, as TXT and JSON. |
| Topics | About 1,000 distinct topic labels chosen by the speakers. The most common are everyday life, food, history, relationships, music, culture, sports and travel. |
| Delivery | Download access within 24 hours of a signed order |
| License | Commercial use under a written license; voice cloning is excluded |
| Price | $35 per hour of audio when you license the full dataset; part of it is quoted on request |
What is in it
| Recording type | Recordings | Hours | Average length |
|---|---|---|---|
| Conversation, open topics: Two American English speakers talk about life, food, history, relationships, music, travel, work and books | 1,517 | 965.3 | 38 min |
The same collection holds another 236 hours in which at least one speaker gives British or another English as their dialect, 130 of them American and British speakers talking to each other. Ask if a mix of accents suits your project.
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.
Getting the data
Listen to the excerpts on the English (US) dataset page and read the machine transcript lines under each one. Need another accent, more speakers or your own recording setup? SpeechData.ai also collects speech data to spec, or you can book a call.
Need something else?
- Call center speech data for call-center style recordings.
- Audio annotation services to transcribe or annotate audio you already have.
- Speech data pricing with published per-hour prices.
Related datasets
Other English datasets from SpeechDataAI on Hugging Face: South African English, Singapore English.
About SpeechData.ai
SpeechData.ai licenses conversational speech datasets in 59 languages and dialects and runs custom speech data collection. It's part of Spirelight. Browse the full catalog.
- Downloads last month
- 68