Synthesizing Buckminster Fuller’s Voice — Research & Recipe

A practical, grounded recipe for the project goal in WHY.md: a synthetic Fuller voice to narrate his texts as audiobooks. This document is research and planning only — it generates no audio and makes no API calls. It is the foundation a later phase executes.

Tool facts below are as of 2026 and volatile — re-verify before building.

1. Voice-data corpus

A faithful clone needs clean, single-speaker reference audio of Fuller. The corpus is unusually good:

In-wiki: see the Everything I Know source article. Everything I Know alone is more than enough reference material — the practical task is selecting and cleaning a few hours of the cleanest segments, not finding audio.

Sources: Internet Archive — Everything I Know, Part 1 · Buckminster Fuller Institute — Everything I Know · “Everything I Know”: 42 Hours of Buckminster Fuller — ArchDaily

2. Tooling — three options + recommendation

The decisive axis is not raw quality (all top tools clone convincingly from minutes of Fuller audio) but license — an audiobook may be distributed, so the cloning engine’s commercial terms matter.

Recommendation: a local, MIT-licensed path built on Chatterbox. Chatterbox is the only top-tier cloning-capable engine with a commercial-safe (MIT) license; it outperformed ElevenLabs in blind listening, self-hosting is dramatically cheaper, and the reference audio stays private. The easiest way to run it is option (d), Voicebox — an MIT desktop app that bundles Chatterbox (plus Kokoro and others) with built-in cloning, a multi-track editor, and post-processing, so most of the §3 pipeline lives in one local tool. Use F5-TTS or XTTS v2 for fast experimentation (their non-commercial licenses are fine for a private proof-of-concept), but ship the distributed audiobook from the permissively-licensed engine (Chatterbox). Avoid ElevenLabs for the primary clone — Professional cloning needs a live session Fuller cannot give, and the Instant clone adds vendor lock-in and ToS exposure for little gain.

Sources: Voicebox (jamiepine/voicebox) · Best Open-Source Alternatives to ElevenLabs (2026) · Local TTS & Voice-Cloning Licenses 2026 · Self-Host Voice Cloning: XTTS-2, F5-TTS, OpenVoice V2 (2026) · Best TTS for Voice Cloning (2026) — CodeSOTA

3. Production pipeline

1. Select & clean reference audio. From Everything I Know, pick 10–30 minutes of the cleanest single-speaker segments (no audience, minimal noise). Loudness-normalize, denoise, and trim silences. Keep a short (~10–30 s) “golden” clip for zero-shot cloning plus a larger set if fine-tuning.
2. Build the voice. Zero-shot: feed the golden clip to Chatterbox/F5-TTS. Optional fine-tune on the larger set for closer timbre and Fuller’s cadence.
3. Prepare the text. Take a book from the repo (e.g. a public-domain or rights-cleared text), split into narratable chunks (paragraph/sentence). Handle Fuller’s coined terms (a pronunciation lexicon — e.g. Dymaxion, tensegrity, omni-), expand abbreviations, and decide how to voice headings, footnotes, and figures.
4. Synthesize chunk-by-chunk, keeping seeds/settings reproducible.
5. Post-process. Concatenate with natural pauses, loudness-normalize to an audiobook target (≈ −19 to −23 LUFS, peaks ≤ −3 dBFS), add chapter markers, export (M4B/MP3).
6. QA — listen for mispronounced coined terms, artifacts, and prosody drift; re-render bad chunks.

4. Quality & evaluation

“Sounds like Bucky” decomposes into three measurable things:

• Intelligibility — can a listener transcribe it accurately? (Spot-check transcription WER.)
• Speaker similarity — does it match Fuller’s timbre? (Speaker-embedding cosine similarity vs. reference, plus human A/B against a real clip.)
• Prosody / pacing — his deliberate, accumulating, slightly halting cadence; harder to score, judged by listening.

Minimal rubric (1–5 each): intelligibility, similarity, prosody, artifact-freeness, coined-term pronunciation. Run a small A/B blind listening test (synthetic vs. a held-out real clip) with a few listeners before committing to a full book. Set a go/no-go bar (e.g. mean ≥ 4 and zero coined-term errors per chapter).

5. Ethics & rights

Synthesizing the voice of a deceased, identifiable person is the part to get right, not an afterthought:

• Rights-holder. The Buckminster Fuller Institute (BFI) is the steward of Fuller’s legacy and hosts Everything I Know; Stanford University Libraries hold the physical archive (the Dymaxion Chronofile). Free access to the recordings is not a grant of voice/likeness reuse rights. Seek explicit permission from BFI (and clarify rights to the underlying recordings) before producing or distributing synthetic-voice audiobooks.
• Disclosure. Any released audio must be clearly labeled as AI-synthesized (title, metadata, and an audible note), never passed off as a genuine recording of Fuller.
• Tool ToS. Honor each engine’s terms on cloning real people (cloud vendors restrict this; respect them). Prefer a permissively-licensed engine you control.
• Spirit. Fuller’s own ethic was to “work always and only for all humanity”; an accessibility audiobook is squarely in that spirit if done transparently and with the estate’s blessing.

Stance: get BFI permission, license the recordings properly, and label the output as synthetic — non-negotiable preconditions for any distribution.

Sources: Buckminster Fuller Institute

6. Phased roadmap + cost

Start at Phase 0 — the permissions gate governs everything downstream. Phase 1 can run in parallel as a private feasibility test, but nothing ships until Phase 0 clears.

Documentation deliverable for work/TODO.md → TODO-6. No audio was generated. Tool/licensing facts are as of 2026 and should be re-verified before execution.