Multilingual Conversations Dataset
Natural conversations for voice AI
People who know each other share context, leave thoughts unfinished, and respond without explicit cues. DataHive’s Multilingual Conversations dataset captures that rhythm: pauses, acknowledgments, interruptions, and spontaneous turns in everyday conversation.
Participants call people they know through Hive Calls and choose their own topics, with no scripts. Recordings come from everyday environments, and participants are paid for qualifying conversation time. The sample supports research on natural conversational voice models and turn-taking dynamics.
What’s in the sample
- Conversations
- 61 conversations · 80 unique speakers
- Total duration
- 19h 31m 47s
- Conversation length
- 2m 34s–58m 57s
- Languages
- English, Hindi, Indonesian, Russian, and Vietnamese
Audio and metadata
The sample provides stereo MP3 audio at 48 kHz and 128 kbps. The caller is on the left channel and the callee on the right, aligned on a shared timeline. Original recordings use separate participant tracks captured through LiveKit as Opus audio in Ogg containers, then converted to MP3 for this sample.
JSONL metadata contains one record per conversation, including topic, language and detected languages, duration, caller and callee country derived from phone numbers, turns per minute, median gap between turns, speech balance, and overlap. Labels and measurements are automatic.
Transcripts are not included in this sample. Audio and metadata are available in the Hugging Face repository’s preview/ folder, with loading instructions in its dataset card.
Quality and selection
Conversations pass QA review and automated checks covering:
- Audio quality: noise, signal-to-noise ratio, clipping, and silence.
- Conversation dynamics: participation by both speakers, turn-taking, overlap, interruptions, and gaps.
- Content: language consistency between tracks, topic, and explicit content.
Conversations are at least two minutes long, and repeat appearances by the same speaker are limited.
Consent and privacy
Participants explicitly consent to recording, sharing with customers, and AI model training. Conversations flagged for personal information are excluded, and published metadata contains no names, phone numbers, or account IDs. Audio and transcript redaction is available on request.
License and access
The sample is released under CC BY 4.0, allowing research and commercial use, including model training, with attribution to DataHive under the license terms. Download the sample and read the full dataset card on Hugging Face.
Custom collections
We collect conversations for voice-model labs by language, accent, and project requirements. Custom deliverables can include separate speaker tracks, transcripts, and quality metadata. On request, we can arrange on-device audio capture before lossy compression, delivered as PCM WAV or FLAC. These are custom collection options, not additional files included in the published sample.
To discuss a custom dataset, use the request form or contact contact@datahive.ai.