FLEURS benchmark excerpts, Google and the FLEURS contributors.
Source: https://huggingface.co/datasets/google/fleurs
Revision: 70bb2e84b976b7e960aa89f1c648e09c59f894dd
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
https://creativecommons.org/licenses/by/4.0/
Dataset paper: https://arxiv.org/abs/2205.12446

These four validation recordings and their transcriptions are excerpts from
FLEURS, not recordings of AnduinOS users. No endorsement by the original
authors is implied. Audio was converted to mono 16 kHz signed 16-bit PCM WAV
with metadata removed; no speech was cut or added. Regression tests may add
deterministic synthetic noise in memory; those are modified samples, not the
original recording conditions. The application never uploads these samples.

Original archives under the pinned revision:
data/en_us/audio/dev.tar.gz
SHA256 2658fda72f199e12676ecac9415094667a4e14e149b146e568ea00b2a2f0954c
data/cmn_hans_cn/audio/dev.tar.gz
SHA256 3bc33212d5974eef7feb04bc4792458d6cd7e14ff10a1a24772f3c45ea87a822
Original transcripts: data/en_us/dev.tsv and data/cmn_hans_cn/dev.tsv
Individual archive member names and converted hashes are in manifest.json.

This small, fixed smoke/regression corpus does not represent all accents,
microphones, languages or noise conditions. Never describe its results as
general speech-recognition accuracy or as Lunar Lake hardware validation.
