Overview
The US-EN Multi-Speaker Conversational Acoustic Corpus provides 24.5 hours of raw, continuous audio capturing authentic multi-party interaction in US English. Featuring 26 unique speakers in discussions with 2–8 participants, the collection supports acoustic research and model development using natural conversational speech.
Recordings preserve overlapping speech, cross-talk, interruptions, spontaneous turn transitions, backchanneling, and disfluencies such as fillers, false starts, and stutters. These dynamics provide material for studying conversational acoustics beyond read or synthetic speech.
TECHNICAL SPECIFICATIONS
• Language and locale: US English (en-US) • Total duration: 24.5 hours • Format: Uncompressed WAV • Sampling rate and bit depth: 44.1 kHz, 16-bit • Session length: 20–30 minutes per continuous file • Participants: 2–8 speakers per discussion • Unique contributors: 26 speakers • Recording environment: Clean, controlled conditions with minimal or no background noise • Post-processing: Raw audio with no destructive filtering • Segmentation: Full continuous sessions without utterance slicing
TOPICS AND DOMAIN COVERAGE
Discussions span 49 distinct topics across six themes: • Healthcare and clinical consultations • Technology, AI, and emerging media • Public policy, ethics, and societal debates • Workplace, education, and career development • Lifestyle, family, and parenting • Culture, media, and entertainment
SESSION-LEVEL METADATA
Each recording is accompanied by its audio file name, total session duration, format verification, participant count, and a roster of unique speaker IDs. Speaker metadata includes gender, age bracket, and a US regional accent or dialect tag. The corpus covers eight major US regional dialects.
DELIVERABLES AND ANNOTATION SCOPE
The base corpus includes continuous audio files and session-level metadata. Text transcripts are not included and are available as an optional add-on tier. Participant rosters identify the speakers present in each session; timestamped speaker turns, utterance boundaries, and text-to-audio alignment are not included.
Speaker consent documentation is available for review.
TARGET USE CASES
Self-supervised speech representation learning: Use natural conversational audio for speech encoder pre-training experiments and acoustic representation learning.
Voice activity detection: Develop and evaluate models using long-form sessions that preserve natural speech, pauses, and overlapping voices.
Accent representation: Explore acoustic embeddings and unsupervised clustering across US regional dialects.
Custom annotation: Apply proprietary transcription, diarization, segmentation, or alignment pipelines to the raw recordings.
Meeting transcription and conversation intelligence: Use the corpus as source audio for model development and evaluation workflows, adding task-specific annotations as needed.
Highlights
- 24.5 hours of continuous US English conversations featuring 26 speakers, 2–8 participants per session, and natural overlap, interruptions, and disfluencies. Delivered as uncompressed 44.1 kHz/16-bit WAV audio with session-level speaker metadata for self-supervised learning, acoustic modeling, and VAD. Transcripts are an optional add-on; timestamped speaker turns are not included.
Details
Features and programs
Financing for AWS Marketplace purchases
Pricing
Custom pricing options
How can we make this page better?
Legal
Content disclaimer
Delivery details
Data as a Service
Subscribe to premium datasets from trusted providers on AWS Marketplace.
Support
Vendor support
Similar products



