simpra/fleurs_xho

FLEURS -- isiXhosa (xh_za)

Re-mirrored from google/fleurs, config xh_za. n-way parallel read speech built on FLORES-101 text -- an evaluation-sized corpus (~19 h), not training scale, but the de facto African-language ASR/TTS benchmark (used in Whisper, MMS, SeamlessM4T, USM papers).

Licence

CC BY 4.0 -- inherited unchanged from the source.

What changed from the source

• audio peak-normalized per clip (see below); sample rate and encoding otherwise unchanged (16000 Hz)
• dropped: rows with missing audio, empty transcript, or exact-duplicate audio payload
• flagged (not dropped) transcripts with a character outside the isiXhosa letter/punctuation/digit set -- FLEURS transcripts legitimately contain digits (e.g. ngo-2011), so this is a QA count, not a filter
• no forced alignment applied -- not applicable, this is already pre-segmented per-utterance audio (see simpra/nchlt_speech_xho's card for when alignment does apply)

Audio quality (measured over the full kept set, not sampled)

Peak-normalized to -3.0 dBFS per clip (single scalar gain per clip, no compression -- relative dynamics within a clip are preserved). Measured before normalizing: google/fleurs xh_za's train split peaks at a median of -20.4 dBFS while test/validation peak at -37.6/-38.2 dBFS -- a ~17 dB gap between splits in the source release that would bias any training/eval mixing them. Mean gain applied per split: {'train': 17.1, 'validation': 35.4, 'test': 35.1}.

	RMS p5	RMS median	RMS p95	RMS spread	peak median
before	-65.3	-57.3	-22.3	43.0	-31.4
after	-36.0	-26.6	-19.0	17.0	-3.0

true clipping before normalizing: 0.34%, checked 4,953/4,953 clips (full corpus)

Result

hours kept: 18.66 h  ·  rows kept: 4,953 / 4,953

split	rows	kept	hours	flag_unknown_chars
train	3,466	3,466	13.34	870
validation	446	446	1.54	61
test	1,041	1,041	3.78	221