speech-aug — a fixed degraded copy of the Sunbird/speech test sets

20,914 utterances / 68.8 hours of African-language speech, across 51 languages and 110 subsets, with telephone-quality and noise degradation baked into the stored audio.

This is a drop-in mirror of Sunbird/speech: same config names, same split name, same columns, same schema. Point an existing evaluation at huwenjie333/speech-aug instead of Sunbird/speech and nothing else has to change.

Why a stored copy rather than augmenting on the fly

Augmenting inside the eval loop draws fresh randomness every run, so two models are never scored on the same audio and a difference of a point or two in WER cannot be separated from the draw. Here one draw is baked in and stored, so every model is scored on byte-identical audio. The trade-off is the usual one: this measures robustness to one fixed realisation of the corruption, not an expectation over the corruption distribution.

Source data

Every row comes from the test split of the corresponding Sunbird/speech config, taken as test[:200] — the first 200 rows in dataset order, not a random sample. 99 of the 110 subsets hit that cap; the rest are smaller test splits taken whole (down to 41 rows).

Text, id, and language are passed through untouched. Only the audio is modified — and duration, which is recomputed from the augmented waveform because the speed op changes clip length.

Subset names follow the source's <language>_<corpus> convention, so the underlying corpora (FLEURS, Common Voice, Waxal, MAK Benchmark, SALT, NextVoices, AfricanVoices, and others) are identifiable from the config name. Consult the source dataset for the provenance, collection method, and terms of each of those corpora.

Augmentation pipeline

Applied in this order to every clip, by analysis/augment_datasets.py --config configs/whisper_augment_af51_test.yaml in SunbirdAI/sunbird-speech. p is the per-clip probability of the op firing.

1. Telephone-bandwidth round trip (p = 0.5)

- set_sample_rate: {rate: 8_000, p: 0.5}   # down
- set_sample_rate: {rate: 16_000}          # and back up, p = 1.0

Half the clips are resampled to 8 kHz and then back to 16 kHz. The down leg discards everything above the 4 kHz Nyquist limit and the up leg cannot restore it, so the result is permanently band-limited — the characteristic dullness of narrowband telephone audio. The other half pass through both steps unchanged (the op is a no-op when the audio is already at the target rate).

Everything is stored at 16 kHz, band-limited or not, so consumers see a uniform sampling rate.

2. Speed-up (p = 0.5)

- augment_audio_speed: {p: 0.5, low: 1.1, high: 1.3}

Half the clips are sped up by a factor drawn uniformly from 1.1–1.3×. This is a WSOLA time stretch (preserve_pitch=True, the default), so the tempo changes but the pitch does not — it reads as a faster talker rather than the tape-speed effect of plain resampling. Sped-up clips are correspondingly shorter, and duration reflects the stored audio.

3. Urban noise over the whole clip (p = 1.0)

- augment_audio_noise:
    min_relative_amplitude: 0.8
    max_relative_amplitude: 1.0
    min_coverage: 1.0
    noise_audio_repo: {path: Sunbird/urban-noise-uganda-61k, name: small, split: train}

Applied to every clip. A noise clip is drawn at random from Sunbird/urban-noise-uganda-61k (config small, split train), tiled if shorter than the speech or randomly cropped if longer, peak-normalised, scaled, and added.

With min_coverage: 1.0 (and max_coverage at its default of 1.0) the noise covers the entire clip, not a segment of it. Its amplitude is uniform(0.8, 1.0) × p99(|speech|), where p99 is the 99th percentile of the absolute speech samples — i.e. noise at roughly 0 dB relative to loud speech. This is deliberately severe: the intent is a hard robustness probe, and word error rates on this data are expected to be far above clean-audio rates.

Note the ordering — noise is added after the resampling round trip, so it is full-band 16 kHz noise even on the clips that were band-limited in step 1.

Clipping

The Audio feature encodes to 16-bit PCM, which hard-clips outside ±1.0, and noise at this amplitude routinely pushes peaks past it. peak_normalize (on by default) rescales any such clip instead, so what is stored is the augmentation the config asked for rather than square-wave distortion.

Reproducibility

Each row's RNG is seeded from (seed=42, row index in the split), so the augmentation of a given utterance depends only on the config, the seed, and its position — not on batch size or worker count. Rerunning the script with this config reproduces this audio.

Schema

Column	Type	Notes
id	string	Source utterance id, unchanged
audio	Audio(sampling_rate=16000)	Augmented, 16-bit PCM wav
text	string	Reference transcript, unchanged
duration	float64	Seconds, recomputed from the augmented audio
language	string	ISO 639-3 code, unchanged

Usage

from datasets import load_dataset

ds = load_dataset("huwenjie333/speech-aug", "lug_fleurs", split="test")
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

Pair a config with the same name in Sunbird/speech to compare clean against degraded on identical utterances, in identical order:

clean = load_dataset("Sunbird/speech", "lug_fleurs", split="test[:200]")
noisy = load_dataset("huwenjie333/speech-aug", "lug_fleurs", split="test")
assert clean["id"] == noisy["id"]   # row order is preserved by the augmentation

Coverage by language

51 languages, 110 subsets, 20,914 clips, 68.8 hours. Language names are ISO 639-3 reference names.

Code	Language	Subsets	Clips	Hours
ach	Acoli	2	296	1.27
afr	Afrikaans	4	729	2.98
aka	Akan	1	200	0.99
amh	Amharic	5	1,000	2.73
bam	Bambara	3	600	0.44
bem	Bemba (Zambia)	2	400	0.80
ber	Berber (collective code, ISO 639-2)	1	200	0.11
cgg	Chiga	1	200	0.55
dag	Dagbani	2	400	1.23
dga	Southern Dagaare	1	200	0.99
eng	English	2	296	0.53
ewe	Ewe	2	400	1.98
fra	French	1	200	0.12
ful	Fulah	3	600	2.37
hau	Hausa	6	1,200	4.00
ibo	Igbo	5	889	2.08
kab	Kabyle	1	200	0.22
kau	Kanuri	1	200	0.36
kik	Kikuyu	1	200	0.32
kin	Kinyarwanda	2	400	0.55
kln	Kalenjin	2	400	0.70
koo	Konzo	1	200	0.87
kpo	Ikposo	1	200	1.03
led	Lendu	1	200	0.89
lgg	Lugbara	1	96	0.20
lin	Lingala	4	800	3.83
lth	Thur	1	171	1.50
lug	Ganda	6	999	3.57
luo	Luo (Kenya and Tanzania)	3	600	1.36
luy	Luyia	1	200	0.48
mlg	Malagasy	1	200	0.92
myx	Masaaba	1	200	1.07
nbl	South Ndebele	1	200	1.19
nya	Chichewa	2	400	0.91
nyn	Nyankole	2	299	1.27
orm	Oromo	3	441	1.32
pcm	Nigerian Pidgin	2	400	0.73
ruc	Ruuli	1	200	1.01
rwm	Amba (Uganda)	1	200	1.14
sna	Shona	3	600	2.83
som	Somali	3	600	1.01
sot	Southern Sotho	2	400	2.34
swa	Swahili (macrolanguage)	2	400	1.02
teo	Teso	1	98	0.16
tsn	Tswana	2	400	1.02
ttj	Tooro	1	200	0.94
wol	Wolof	3	600	1.94
xho	Xhosa	3	600	2.64
xog	Soga	1	200	1.17
yor	Yoruba	5	1,000	1.94
zul	Zulu	3	600	3.15
	Total	110	20,914	68.8

Coverage by subset (110 configs)

Config	Language	Clips	Hours
ach_salt	ach	96	0.14
ach_waxal	ach	200	1.13
afr_afrikaans30s	afr	200	1.52
afr_commonvoice	afr	129	0.21
afr_fleurs	afr	200	0.63
afr_makbenchmark	afr	200	0.63
aka_waxal	aka	200	0.99
amh_commonvoice	amh	200	0.33
amh_kyagaba	amh	200	0.37
amh_makbenchmark	amh	200	0.56
amh_shunyalabs	amh	200	0.58
amh_waxal	amh	200	0.90
bam_makbenchmark	bam	200	0.16
bam_robotsbambari	bam	200	0.17
bam_robotsmali	bam	200	0.11
bem_csikasote	bem	200	0.40
bem_makbenchmark	bem	200	0.40
ber_tutlay	ber	200	0.11
cgg_commonvoice	cgg	200	0.55
dag_commonvoice	dag	200	0.23
dag_waxal	dag	200	1.00
dga_waxal	dga	200	0.99
eng_intronhealth	eng	200	0.41
eng_salt	eng	96	0.12
ewe_makbenchmark	ewe	200	1.01
ewe_waxal	ewe	200	0.98
fra_gigant	fra	200	0.12
ful_fleurs	ful	200	0.70
ful_makbenchmark	ful	200	0.70
ful_waxal	ful	200	0.98
hau_africanvoices	hau	200	2.05
hau_clearglobal	hau	200	0.37
hau_commonvoice	hau	200	0.25
hau_fleurs	hau	200	1.02
hau_makbenchmark	hau	200	0.15
hau_naijavoices	hau	200	0.16
ibo_africanvoices	ibo	200	0.68
ibo_commonvoice	ibo	89	0.13
ibo_fleurs	ibo	200	0.88
ibo_makbenchmark	ibo	200	0.19
ibo_naijavoices	ibo	200	0.21
kab_commonvoice	kab	200	0.22
kau_clearglobal	kau	200	0.36
kik_anvkekikuyu	kik	200	0.32
kin_commonvoice	kin	200	0.27
kin_makbenchmark	kin	200	0.28
kln_ankekalenjin	kln	200	0.39
kln_commonvoice	kln	200	0.32
koo_commonvoice	koo	200	0.87
kpo_waxal	kpo	200	1.03
led_commonvoice	led	200	0.89
lgg_salt	lgg	96	0.20
lin_fleurs	lin	200	0.98
lin_kasuletrev	lin	200	0.93
lin_shunyalabs	lin	200	0.98
lin_waxal	lin	200	0.94
lth_commonvoice	lth	171	1.50
lug_commonvoice	lug	200	0.31
lug_fleurs	lug	200	0.87
lug_makbenchmark	lug	200	0.86
lug_makerereradio	lug	100	0.17
lug_salt	lug	99	0.15
lug_waxal	lug	200	1.20
luo_anvkeluo	luo	200	0.43
luo_commonvoice	luo	200	0.23
luo_fleurs	luo	200	0.70
luy_digitaldivide	luy	200	0.48
mlg_waxal	mlg	200	0.92
myx_waxal	myx	200	1.07
nbl_nextvoices	nbl	200	1.19
nya_fleurs	nya	200	0.83
nya_michsethowusu	nya	200	0.09
nyn_salt	nyn	99	0.16
nyn_waxal	nyn	200	1.11
orm_fleurs	orm	41	0.12
orm_makbenchmark	orm	200	0.37
orm_waxal	orm	200	0.82
pcm_africanvoices	pcm	200	0.46
pcm_commonvoice	pcm	200	0.26
ruc_commonvoice	ruc	200	1.01
rwm_commonvoice	rwm	200	1.14
sna_fleurs	sna	200	0.75
sna_makbenchmark	sna	200	1.04
sna_waxal	sna	200	1.05
som_anvkesomali	som	200	0.22
som_fleurs	som	200	0.70
som_skydheere	som	200	0.09
sot_fleurs	sot	200	0.94
sot_nextvoices	sot	200	1.41
swa_commonvoice	swa	200	0.30
swa_fleurs	swa	200	0.72
teo_salt	teo	98	0.16
tsn_commonvoice	tsn	200	0.22
tsn_nextvoices	tsn	200	0.80
ttj_commonvoice	ttj	200	0.94
wol_fleurs	wol	200	0.88
wol_kallama	wol	200	0.21
wol_makbenchmark	wol	200	0.85
xho_fleurs	xho	200	0.65
xho_makbenchmark	xho	200	0.66
xho_nextvoices	xho	200	1.33
xog_waxal	xog	200	1.17
yor_africanvoices	yor	200	0.45
yor_commonvoice	yor	200	0.33
yor_fleurs	yor	200	0.80
yor_makbenchmark	yor	200	0.18
yor_naijavoices	yor	200	0.18
zul_fleurs	zul	200	0.84
zul_makbenchmark	zul	200	0.84
zul_nextvoices	zul	200	1.46

Intended use and limitations

Intended for evaluating ASR robustness to noisy, narrowband, fast speech — not for training, and not as a measure of clean-audio accuracy.

• Not a random sample. test[:200] takes the first rows in dataset order. If a source split is grouped by speaker or session, the cap inherits that grouping, and per-subset numbers here need not match the full test split.
• One fixed draw. Results characterise robustness to this particular realisation of the corruption, not an expectation over it.
• Severe by design. ~0 dB noise over the full clip is harsher than most deployment conditions; treat it as a stress test, not a field estimate.
• Unequal subset sizes. Subsets range from 41 to 200 clips, so a macro-average over subsets weights small ones heavily. Aggregate deliberately.
• Inherited properties. Transcript quality, speaker demographics, dialect coverage, and licensing all come from the source corpora and are unchanged here. Sunbird/speech declares no license; the terms of the underlying corpora apply, and this derived copy claims no additional rights.

Provenance

Source	Sunbird/speech, test splits, test[:200]
Noise source	Sunbird/urban-noise-uganda-61k, config small, split train
Script	analysis/augment_datasets.py
Config	configs/whisper_augment_af51_test.yaml
Repo	SunbirdAI/sunbird-speech
Seed	42