Whisper Tiny English and multilingual GGML
Owner-requested compact Whisper tier. Tiny is a 2022 architecture, a labelled resource-tier exception rather than a latest-generation accuracy claim. Both exact variants passed short English synthetic speech, silence and recovery via the Windows public SDK. Tiny is multilingual (99 language codes); Tiny.en is English-only. Larger Whisper variants have higher published accuracy. Mobile, every-language and corpus quality remain unqualified.
Primary selection evidence. Publisher-reported quality is separate from independent Wekala measurements. This catalogue is not a claim that every voice/language has equal quality.
Original weight/data notices and exact source provenance are preserved. The model license permits commercial use subject to its terms; it grants no rights to third-party identities, reference recordings or application content. Executable engines, phonemizers, watermark libraries and their dependency/data rights require separate records.
All assets remain consumer supplied. This repository contains declarative model/tokenizer/voice data and notices only. It contains no Wekala SDK, playground, fixture, runtime binary, Python/Node code, generated speech, protected content or credentials. Nothing here selects a fallback or installs an engine. Configuration files are data for a separately qualified bundled implementation.
source-manifest.json preserves immutable source revisions, hashes and structural inspection results. files.json binds the exact mirror. Native inference, all-language intelligibility, long-form stability, sustained mobile/browser performance, physical audio and production release acceptance are separate gates.