DariusTheGeek/waxal-ttia-gallery
Waxal TTIA enrolment gallery Data assets for the Test-Time Idiolect Adaptation (TTIA) Lingala lane of the waxal-asr-solution pipeline (Google Waxal ASR Challenge on Zindi). Fetched into place by models/download_assets.py in that repository; every file is verified there against a recorded SHA-256. File Size Contents enrollment.npz 998 MB MMS-1B hidden-state voice vectors for 21,566 enrolment clips (layers 4, 6, 8, 12, 16; the pipeline reads 4 and 8), plus their clip… See the full description on the dataset page: https://huggingface.co/datasets/DariusTheGeek/waxal-ttia-gallery.
Waxal TTIA enrolment gallery
Data assets for the Test-Time Idiolect Adaptation (TTIA) Lingala lane of the waxal-asr-solution pipeline (Google Waxal ASR Challenge on Zindi). Fetched into place by models/download_assets.py in that repository; every file is verified there against a recorded SHA-256.
Provenance and licence
Derived from google/WaxalNLP (Lingala labelled training audio and transcripts, and clips from the unlabelled release), © the WaxalNLP authors, licensed CC-BY-SA-4.0 / CC-BY-4.0. These derivatives are published under CC-BY-SA-4.0 with attribution to google/WaxalNLP, as ShareAlike requires. No evaluation/test audio or transcripts are included.
The files are reproducible from google/WaxalNLP with inference/ttia/build_enrollment.py, embed.py and merge.py in the solution repository; this upload exists so a reviewer can run the pipeline without rebuilding them (a ~21,500-clip GPU embedding pass).