Leobuildsthings/mentara-conversations

Mentara live voice-persona conversations (free sample)

Real people talking live, by voice, with a real-time AI persona: interruptions, half-finished lines, slang, banter and mishearings that scripted and synthetic chat doesn't have. Prepared by Mentara Purifier: consented, de-identified, cleaned and quality-checked.

This repo is a 174-row sample. The full dataset is 1,392 rows from 46 calls and growing every month.

• See the persona live: 100+ YouTube videos of it talking with real people
• The product the data comes from: Mentara Persona
• Buy the full dataset: request it here

Full dataset pricing

• $0.45 per row, $1,000 minimum (the current 1,392 rows: about $1,000)
• $0.30 per row above 10,000 rows
• $300/month for each new monthly release (the dataset keeps growing)
• Non-exclusive licence; exclusive on request

Available packs

The same cleaned rows, split by the persona's speaking mode, so you buy only the styles you can use.

Pack	Modes	Rows	Price	For
Full pack	every mode	1,392	$1,000	companion, character, entertainment and gaming apps
Clean pack	none, comfort	783	$1,000	enterprise, customer-facing and voice-agent teams
Companion pack	none, comfort, flirty	1,220	$1,000	companion and romance apps that don't want the edgier modes

Modes:

• none: normal conversation, the persona's everyday voice
• comfort: supportive and warm
• flirty: playful and romantic, nothing explicit
• troll: teasing, comebacks, playful insults
• unhinged: chaotic, absurd humour; not cruel

Format

JSON Lines, chat format. One caller line and the persona's reply per row:

{"id": "s0001", "call": 1,
  "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
  "meta": {"persona": "...", "mode": "troll", "group_call": false, "split": "train",
           "stt_confidence": -0.02, "consent": {"status": "consented", ...}}}

• stt_confidence: speech-to-text confidence for the caller line (0 = certain).
• group_call: more than one caller; speakers are marked [Speaker N].
• split: train/eval, held out by whole call.
• consent: on every row.

Full dataset at a glance

rows	1,392
calls	46
modes	none 757, flirty 437, troll 172, comfort 26
group-call rows	19

Consent

Every person recorded agreed directly to recording, AI training and resale of de-identified conversation data, and can withdraw at any time (withdrawals are removed from future releases). Every row carries the consent stamp.

How it was cleaned

De-identified (emails, URLs, handles, numbers by pattern; names and places by an NER model); likely mishearings, canned replies, duplicates, over-repeated replies and slurs removed; the app's speech effects undone so the text is what the persona wrote.

This sample is stricter than the full set: no explicit lines, no rows with redacted names, confident transcripts only. The full set includes adult, sometimes crude, language (flirty, troll and unhinged modes).

Known limits

• Few speakers: most calls come from a small group of regular callers.
• English, US-centric; gaming and casual-chat topics.
• The persona side is AI-generated.
• Text only: no audio or voiceprints are included or sold.

Licence

The sample is CC BY-NC 4.0: free to inspect and test, not for commercial training. The full dataset is sold under a commercial licence (training and evaluating AI models; no re-identification, no resale).

_{Sample built 2026-11-04 from release 2026-11-03.}