Datasets:
license: cc-by-4.0
language:
- hi
pretty_name: DataHive — Hindi Conversations (Sample)
task_categories:
- automatic-speech-recognition
- audio-classification
- voice-activity-detection
tags:
- conversational-speech
- turn-taking
- stereo
- dialogue
size_categories:
- n<1K
DataHive — Hindi Conversations
Unscripted conversations in Hindi between people who already know each other. Shared to support research on natural conversational voice models.
Why These Conversations
Natural conversation depends on timing: when to respond, brief acknowledgments, pauses, and interruptions. People who know each other share context, leave thoughts unfinished, and switch turns without explicit signals. These recordings capture that.
How We Collect
Participants call people they know through the Hive Calls app (Google Play, App Store). They choose their own topics, with no scripts. Conversations take place in everyday environments and are recorded through LiveKit, with each participant saved separately using Opus compression in an Ogg container. These recordings are converted to stereo MP3 for this sample. Participants are paid for qualifying conversation time.
What’s in This Sample
| Detail | Value |
|---|---|
| Language | Hindi (hi) |
| Conversations | 22 |
| Speakers | 35 unique speakers |
| Total duration | 4h 13m 34s |
| Conversation length | 5m 4s–21m 59s |
| Audio | Stereo MP3 · 48 kHz · 128 kbps |
| Channels | Caller left, callee right, on a shared timeline |
| Metadata | JSONL, one record per conversation |
| License | CC BY 4.0 |
Metadata per conversation:
- topic
language: conversation language code (hi)languages: detected language codes from the original export- duration
- caller and callee country (from phone number)
- turns per minute
- median gap between turns
- speech balance
- overlap
Transcripts are not included.
Files
README.md
preview/metadata.jsonl
preview/hi_000001.mp3
preview/hi_000002.mp3
...
Install the Hugging Face audio dependencies, then load the dataset from the extracted folder:
python -m pip install "datasets[audio]"
from datasets import load_dataset
ds = load_dataset("audiofolder", data_dir="preview")
Quality and Selection
Every conversation passes QA review and automated checks for:
- Audio: noise, signal-to-noise ratio, clipping, silence.
- Dynamics: both speakers participate, turn-taking, overlap, interruptions, gaps.
- Content: the same language on both tracks, topic, explicit content.
Conversations are 5–30 minutes long. Labels and measurements are automatic.
How This Sample Was Drawn
The conversations were drawn at random with a recorded seed, not hand-picked or optimised.
- Pool: every Hindi conversation that passes the checks above and is 5–30 minutes long.
- Order: the pool is sorted by
sha256("<salt>:<call id>"). - Draw: conversations are taken in that order, skipping any that would break a rule:
- a pair of speakers appears in at most one conversation;
- a speaker appears in at most 2 conversations;
- the sample stays under 5 hours.
- Check: the sample is kept only if it is at least 3 hours long and no conversation is more than 15% of it.
- If a check fails, the next salt (1, 2, 3, …) is tried. The first salt that passes is published.
| Detail | Value |
|---|---|
| Salt | 1 |
Consent and Privacy
Participants explicitly consent to recording, sharing with customers, and AI model training. Conversations flagged for personal information are excluded. Metadata contains no names, phone numbers, or account IDs. Audio and transcript redaction is available on request.
Custom Collections
We collect conversations for voice-model labs by language, accent, and requirements. Deliverables can include separate speaker tracks, transcripts, and quality metadata. On request, we can arrange collection of audio captured on-device before lossy compression, delivered as PCM WAV or FLAC. Contact contact@datahive.ai.
License
Released under CC BY 4.0. Free for research and commercial use, including model training, with attribution to DataHive.