CoolFace
Datasetpublic

akshan-main/glossolalia-inputs

Glossolalia Dial Inputs Input data for the Glossolalia Dial: one dial that grades a typed sentence from clean speech to wordless tongues in the same voice. These are the building blocks the pipeline runs on. The 30k-clip training corpus is regenerated from them with the scripts in the code repo, so it is not stored here. Contents File What it is sentences.txt Source English sentences. The dial learns to dissolve sentences like these. The corruption… See the full description on the dataset page: https://huggingface.co/datasets/akshan-main/glossolalia-inputs.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face
0likes29downloads
Dataset Card

Glossolalia Dial Inputs

Input data for the Glossolalia Dial: one dial that grades a typed sentence from clean speech to wordless tongues in the same voice. These are the building blocks the pipeline runs on. The 30k-clip training corpus is regenerated from them with the scripts in the code repo, so it is not stored here.

Contents

FileWhat it is
sentences.txtSource English sentences. The dial learns to dissolve sentences like these. The corruption pipeline reads them; the model never sees the corrupted text.
voices/v1.wav ... voices/v9.wavNine reference voice clips for zero-shot cloning (F5-TTS clones whichever voice you pick).
voices/v1.txt ... voices/v9.txtTranscript of each reference clip, which improves clone quality.
cmudict.dictCMU Pronouncing Dictionary. Ghost mode looks up each word's phonemes here to find close-sounding real words.
phoneme_lm.npzPhoneme unigram and bigram model fit from CMUdict, used by the Tongues-mode phoneme corruption.

How it is used

  • —Tongues mode corrupts a sentence's phonemes (using phoneme_lm.npz) at rising rates, has base F5-TTS read each version, and trains a LoRA plus a scalar dial to reproduce that slide from the clean sentence alone.
  • —Ghost mode runs live: for each word it searches cmudict.dict for real words within a phonetic-feature distance, then reranks with a small language model into a plausible mishearing.

Sources and licensing

  • —Sentences: public-domain text (Project Gutenberg) plus LibriSpeech transcripts (CC-BY-4.0).
  • —Voices: reference clips from openly-licensed read speech (LibriVox public domain and LibriTTS-R, CC-BY-4.0).
  • —CMUdict: CMU Pronouncing Dictionary, BSD-2-Clause.
  • —Phoneme LM: derived from CMUdict.

Attribute LibriSpeech and LibriTTS-R per CC-BY-4.0. No scraped or cloud-sourced data; the set is rights-clean and the pipeline reproduces from it end to end.