Spaces:

---

otoearth

/

README

like 0

Running

Data Collection Protocol – Draft Proposal (v0.1)

Summary
Proposed spec for collecting spontaneous two-speaker conversations: audio parameters, per-speaker channeling, metadata, random topic prompts, and an in-recording self-redaction feature.

A. Audio & Channeling

• Format: PCM WAV, 48 kHz, 16-bit
• Channeling: Dual-mono (Speaker A / Speaker B) per session; optional mixed-down reference track
• Environment: Everyday ambient noise allowed; copyrighted audio (music/TV/podcasts) is not allowed

B. Session Design

• Session length: 25 minutes (fixed) per session
• Chunking: Not used. If 5-minute chunking becomes operationally necessary, we will open a separate discussion before adopting it
• Starter prompts: Show 1–3 random topics at session start (e.g., favorite foods, countries visited, recent books)
• Task-oriented dialogs: Out of scope for now; may be added if community consensus emerges (e.g., trip planning, brainstorming, tongue-twisters, word-chain)

C. File Naming (TBD)

• Example (session-level files):
  YYYYMMDD_sessionID_speaker{A|B}.wav
  Companion JSON (session-level): YYYYMMDD_sessionID.meta.json
  Do not place any personally identifying information in file names.

D. Metadata (per participant / session)
All profile fields are opt-out (per-field). Items may be masked or withheld at release time based on re-identification risk.

• speaker_id — stable pseudonymous public ID per participant (e.g., spk_6Z4G3Y9Q). A separate internal ID is maintained privately for withdrawal/compliance workflows
• age (e.g., 32)
• gender (self-identified)
• nationality
• birth_country and birth_state/prefecture
• accent (e.g., Japanese English)
• first_language, second_language
• education_level (e.g., bachelor’s degree)
• MBTI (optional)
• occupation (e.g., entrepreneur)
• residence_country
• interests (e.g., AI, Crypto)
• device/OS/microphone (technical metadata)
• network/latency logs (quality metadata)

E. In-Recording Privacy Control

• Self-redaction: During recording, a speaker can delete the most recent 10 seconds of their own speech, which is removed on-device and not uploaded
• UI hint: A “rewind 10s → delete” control with a confirmation dialog

Open Questions

1. Naming convention: Start with YYYYMMDD_sessionID_speakerX or include short codes for language/locale/topic?
2. Chunking trigger: Under what operational conditions (if any) should we introduce 5-minute chunking (e.g., moderation backlog thresholds)?
3. High-risk fields: For items like alma mater, should we adopt “collect-yes / public-by-request or withheld-by-default”?
4. Self-redaction duration: Keep 10s fixed, or offer selectable 5/10/20s?
5. Non-dual-mono submissions: Should we accept WebRTC-style separated streams as an alternative to dual-mono?