build-small-hackathon/glossolalia-dial

Glossolalia Dial

A single dial that grades a typed lyric into dreamy territory in two distinct phonotactic paths:

• —Ghost mode — the lyric is rewritten as a sequence of real English words that phonetically resemble the source. The substitution is fully deterministic (PanPhon feature-edit distance over CMUdict, syllable-count + primary-stress matched) and reranked by DistilGPT-2 beam search for semantic coherence. F5-TTS base reads it.
• —Tongues mode — the clean lyric goes into F5-TTS + a fine-tuned LoRA + a learned scalar conditioner (LevelEmbed at AdaLN side). The LoRA produces graded glossolalic audio in the chosen voice — invented pseudowords, sonorant-leaning palette.

Both modes ride F5-TTS for voice cloning. Off-the-Grid: no cloud APIs.

**Source code** · **Fine-tuned LoRA** · **Data inputs**

How it works

Ghost mode:    lyric → mondegreen.substitute() → real-English ghost → F5-TTS base → audio
Tongues mode:  lyric → F5-TTS + v8 LoRA + LevelEmbed(dial scalar) → graded glossolalic audio

Honest provenance

Every load-bearing design decision lives in DECISIONS.md. A citation audit refuted 6/10 originally-cited sources; the remaining provenance is verified or labeled as design heuristic with no false attribution.

Badges targeted

• —🎯 Well-Tuned — published fine-tuned LoRA at akshan-main/glossolalia-dial-lora
• —🔌 Off the Grid — no cloud APIs; all inference local
• —📓 Field Notes — BLOG.md walks the audit + iteration trail