Oswald compare

Speakers

V2 and V3 never trained on held-out speakers

Speaker

Gender

Department

Text contains

Clips: click a row to add it to the selection and hear it

Selected clips (# from the table)

Random clips

500 clips on all four models take about 12 GPU minutes

1 500

Last clicked clip

Scores (each model as soon as it has finished)

All outputs (CSV)

Kreyòl audio (up to 2 min)

What was said (optional, for scores)

Scores

Datamio clips. jsbeaudry/datamio-kreyol-voices: approved phone recordings of sentences read aloud. The 14 speakers V2 held out (10 test, 4 validation) stayed out of V3's training too, so only they are fair ground for comparing V2 and V3. V1 never trained on Datamio at all. Training speakers' clips were heard in training by V2 (takes to Oct 15) and V3.

Scores. Word and character error rate after reading digits back as words, lowercasing and dropping punctuation and accents, as in the V3 evaluation. Numbers whole counts a number right only when all its words come out. In each clip's card, scored as shows the output in that form: struck red = missed, orange = extra or wrong.

Models. Oswald V1: jsbeaudry/Oswald-ASR-0.6-V1; Oswald V2: jsbeaudry/Oswald-ASR-0.6-V2; Oswald V3: jsbeaudry/Oswald-ASR-0.6-V3; Oswald V3 no-JW: jsbeaudry/Oswald-ASR-0.6-V3-nojw. V1/V2 decode with beam 4 and drop phrases repeated more than three times (as in the Trio Space); V3 decodes greedily (TDT). V3 with JW is private by the jw.org terms; the no-JW arm has no jw.org data.