Hugging Face's logo

Datasets:
Sunbird
/
speech-benchmark

Tasks:
Automatic Speech Recognition
Modalities:
Audio
Text
Formats:
parquet
Languages:
Acoli
Afrikaans
Akan
Size:
10K - 100K
Tags:
speech
asr
african-languages
benchmark
multilingual
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Sunbird Speech Benchmark

A multilingual automatic speech recognition (ASR) test set for 51 languages spoken in Africa: 20,213 utterances (64.8 hours) drawn from 110 subsets of 24 source corpora.

This is the exact set of clips used to evaluate Sunbird/SunflowerASR-51-african-languages (SunflowerASR) and the systems it was compared against (Gemini, GPT-4o Transcribe, Omnilingual ASR). Freezing the clips in one repository means every model can be scored on identical audio and references, and the per-language row counts reproduce the n_examples of the published results.

Languages: Acholi (ach), Afrikaans (afr), Akan (aka), Amharic (amh), Ateso (teo), Bambara (bam), Bemba (bem), Berber (ber), Chichewa (nya), Dagaare (dga), Dagbani (dag), English (eng), Ewe (ewe), French (fra), Fulani (ful), Hausa (hau), Igbo (ibo), Ikposo (kpo), Kabyle (kab), Kalenjin (kln), Kanuri (kau), Kikuyu (kik), Kinyarwanda (kin), Kwamba (rwm), Lendu (led), Lingala (lin), Luganda (lug), Lugbara (lgg), Luhya (luy), Lumasaba (myx), Luo (luo), Lusoga (xog), Malagasy (mlg), Ndebele (nbl), Nigerian Pidgin (pcm), Oromo (orm), Rukiga (cgg), Rukonjo (koo), Runyankole (nyn), Ruruuli (ruc), Rutooro (ttj), Shona (sna), Somali (som), Sotho (sot), Swahili (swa), Thur (lth), Tswana (tsn), Wolof (wol), Xhosa (xho), Yoruba (yor), Zulu (zul).

Quick start 

from datasets import load_dataset, Audio

# Everything (20,213 rows)
ds = load_dataset("Sunbird/speech-benchmark", split="test")

# One language, e.g. Luganda (all of its source subsets)
lug = load_dataset("Sunbird/speech-benchmark", "lug", split="test")

# One source subset, named as in Sunbird/speech
lug_salt = load_dataset("Sunbird/speech-benchmark", "lug_salt", split="test")

# Audio is stored in its original encoding and resampled to 16 kHz on access
sample = lug[0]
sample["audio"]["array"], sample["audio"]["sampling_rate"], sample["text"]

Access requires a Hugging Face account with permission to this repository; authenticate with hf auth login (or set HF_TOKEN) before loading.

Configurations

Config Contents
all (default) All 20,213 rows
<language code> e.g. lug, swa, yor All subsets of one language (51 configs)
<code>_<source> e.g. lug_salt, hau_fleurs One source subset (110 configs)

Every configuration has a single test split.

Data fields

Field Type Description
id string Utterance id, carried over from Sunbird/speech
audio Audio Original encoded audio bytes (OGG/OPUS), decoded and resampled to 16 kHz when accessed
text string Reference transcript, unnormalized, as in the source
duration float64 Clip duration in seconds
language string ISO 639-3 code (the language token forced at decoding time)
language_name string English language name
subset string Source subset name, <code>_<source>
source string Name of the upstream corpus

Dataset construction

Rows were selected from the held-out test splits of Sunbird/speech (revision 4edec52), a harmonised collection of public and community-collected African speech corpora (every subset has its own train/dev/test split), replaying the evaluation pipeline exactly:

  1. For each of the 110 subsets of the 51 languages, take the first 200 rows of the test split in storage order (test[:200], no shuffling).
  2. Drop clips longer than 30 s, comparing durations rounded to 10 ms (698 clips). The rounding keeps one boundary clip, nextvoices_sot_8b28b599 (30.004 s), which is scored in the published SunflowerASR results.
  3. Drop clips whose reference transcript is empty after text normalisation (3 clips), since they cannot be scored.

52 of the 110 subsets contribute the full 200 clips; the rest have smaller test splits or lost clips to the length filter. Audio and transcripts are otherwise unchanged.

Statistics

By language

Language Code Subsets Examples Hours Mean dur. (s) Sources
Acholi ach 2 270 1.13 15.1 Google WAXAL; Sunbird SALT
Afrikaans afr 4 729 3.27 16.1 ASR Africa Data Efficiency Benchmark; Afrikaans-30s; Google FLEURS; Mozilla Common Voice
Akan aka 1 200 1.08 19.5 Google WAXAL
Amharic amh 5 1,000 2.99 10.8 ASR Africa Data Efficiency Benchmark; Google WAXAL; KYAGABA; Mozilla Common Voice; Shunya Labs
Ateso teo 1 98 0.18 6.5 Sunbird SALT
Bambara bam 3 600 0.48 2.9 ASR Africa Data Efficiency Benchmark; RobotsMali Bambara ASR
Bemba bem 2 400 0.87 7.9 ASR Africa Data Efficiency Benchmark; BembaSpeech
Berber ber 1 200 0.12 2.1 TutlaytAI Amazigh ASR
Chichewa nya 2 399 0.99 8.9 Google FLEURS; michsethowusu
Dagaare dga 1 200 1.09 19.6 Google WAXAL
Dagbani dag 2 400 1.35 12.1 Google WAXAL; Mozilla Common Voice
English eng 2 295 0.58 7.0 AfriSpeech-200 (Intron Health); Sunbird SALT
Ewe ewe 2 400 2.17 19.6 ASR Africa Data Efficiency Benchmark; Google WAXAL
French fra 1 200 0.13 2.4 African-Accented French
Fulani ful 3 594 2.55 15.5 ASR Africa Data Efficiency Benchmark; Google FLEURS; Google WAXAL
Hausa hau 6 1,130 2.19 7.0 ASR Africa Data Efficiency Benchmark; African Voices; CLEAR Global; Google FLEURS; Mozilla Common Voice; NaijaVoices
Igbo ibo 5 860 1.79 7.5 ASR Africa Data Efficiency Benchmark; African Voices; Google FLEURS; Mozilla Common Voice; NaijaVoices
Ikposo kpo 1 200 1.13 20.3 Google WAXAL
Kabyle kab 1 200 0.24 4.3 Mozilla Common Voice
Kalenjin kln 2 400 0.77 6.9 African Next Voices (Kenya); Mozilla Common Voice
Kanuri kau 1 200 0.40 7.1 CLEAR Global
Kikuyu kik 1 199 0.34 6.1 African Next Voices (Kenya)
Kinyarwanda kin 2 400 0.61 5.5 ASR Africa Data Efficiency Benchmark; Mozilla Common Voice
Kwamba rwm 1 199 1.24 22.5 Mozilla Common Voice
Lendu led 1 199 0.97 17.5 Mozilla Common Voice
Lingala lin 4 767 3.89 18.3 Google FLEURS; Google WAXAL; KasuleTrevor Lingala; Shunya Labs
Luganda lug 6 950 3.41 12.9 ASR Africa Data Efficiency Benchmark; Google FLEURS; Google WAXAL; Makerere Radio Speech; Mozilla Common Voice; Sunbird SALT
Lugbara lgg 1 96 0.22 8.3 Sunbird SALT
Luhya luy 1 200 0.52 9.4 Digital Divide Data
Lumasaba myx 1 180 0.98 19.5 Google WAXAL
Luo luo 3 596 1.45 8.7 African Next Voices (Kenya); Google FLEURS; Mozilla Common Voice
Lusoga xog 1 177 1.04 21.2 Google WAXAL
Malagasy mlg 1 199 1.00 18.1 Google WAXAL
Ndebele nbl 1 144 0.64 16.1 African Next Voices (Southern Africa)
Nigerian Pidgin pcm 2 397 0.71 6.4 African Voices; Mozilla Common Voice
Oromo orm 3 439 1.42 11.6 ASR Africa Data Efficiency Benchmark; Google FLEURS; Google WAXAL
Rukiga cgg 1 200 0.60 10.9 Mozilla Common Voice
Rukonjo koo 1 200 0.96 17.3 Mozilla Common Voice
Runyankole nyn 2 280 1.20 15.4 Google WAXAL; Sunbird SALT
Ruruuli ruc 1 170 0.78 16.5 Mozilla Common Voice
Rutooro ttj 1 199 1.02 18.4 Mozilla Common Voice
Shona sna 3 594 3.05 18.5 ASR Africa Data Efficiency Benchmark; Google FLEURS; Google WAXAL
Somali som 3 596 1.07 6.5 African Next Voices (Kenya); Google FLEURS; Soomali ASR
Sotho sot 2 321 1.44 16.1 African Next Voices (Southern Africa); Google FLEURS
Swahili swa 2 400 1.12 10.1 Google FLEURS; Mozilla Common Voice
Thur lth 1 167 0.67 14.5 Mozilla Common Voice
Tswana tsn 2 369 0.78 7.6 African Next Voices (Southern Africa); Mozilla Common Voice
Wolof wol 3 571 1.83 11.5 ASR Africa Data Efficiency Benchmark; Google FLEURS; Kallaama
Xhosa xho 3 527 2.02 13.8 ASR Africa Data Efficiency Benchmark; African Next Voices (Southern Africa); Google FLEURS
Yoruba yor 5 995 2.03 7.4 ASR Africa Data Efficiency Benchmark; African Voices; Google FLEURS; Mozilla Common Voice; NaijaVoices
Zulu zul 3 507 2.31 16.4 ASR Africa Data Efficiency Benchmark; African Next Voices (Southern Africa); Google FLEURS
Total 110 20,213 64.82 11.5 24 sources

By source dataset

Source dataset Languages Subsets Examples Hours License (upstream)
Mozilla Common Voice Afrikaans, Amharic, Dagbani, Hausa, Igbo, Kabyle, Kalenjin, Kinyarwanda, Kwamba, Lendu, Luganda, Luo, Nigerian Pidgin, Rukiga, Rukonjo, Ruruuli, Rutooro, Swahili, Thur, Tswana, Yoruba 21 3,952 10.19 CC0-1.0
ASR Africa Data Efficiency Benchmark Afrikaans, Amharic, Bambara, Bemba, Ewe, Fulani, Hausa, Igbo, Kinyarwanda, Luganda, Oromo, Shona, Wolof, Xhosa, Yoruba, Zulu 16 3,166 9.42 See source
Google FLEURS Afrikaans, Chichewa, Fulani, Hausa, Igbo, Lingala, Luganda, Luo, Oromo, Shona, Somali, Sotho, Swahili, Wolof, Xhosa, Yoruba, Zulu 17 3,128 13.16 CC-BY-4.0
Google WAXAL Acholi, Akan, Amharic, Dagaare, Dagbani, Ewe, Fulani, Ikposo, Lingala, Luganda, Lumasaba, Lusoga, Malagasy, Oromo, Runyankole, Shona 16 3,053 16.35 CC-BY-4.0 / CC-BY-SA-4.0
African Next Voices (Kenya) Kalenjin, Kikuyu, Luo, Somali 4 797 1.45 See source
African Voices Hausa, Igbo, Nigerian Pidgin, Yoruba 4 741 1.59 See source
African Next Voices (Southern Africa) Ndebele, Sotho, Tswana, Xhosa, Zulu 5 709 3.02 CC-BY-4.0
NaijaVoices Hausa, Igbo, Yoruba 3 600 0.60 CC-BY-NC-SA-4.0
Sunbird SALT Acholi, Ateso, English, Luganda, Lugbara, Runyankole 6 584 1.02 CC-BY-SA-4.0
RobotsMali Bambara ASR Bambara 2 400 0.30 See source
CLEAR Global Hausa, Kanuri 2 400 0.80 See source
Shunya Labs Amharic, Lingala 2 387 1.59 See source
Digital Divide Data Luhya 1 200 0.52 See source
BembaSpeech Bemba 1 200 0.44 See source
African-Accented French French 1 200 0.13 CC
Afrikaans-30s Afrikaans 1 200 1.67 CC-BY-4.0
TutlaytAI Amazigh ASR Berber 1 200 0.12 See source
Soomali ASR Somali 1 200 0.09 CC-BY-4.0
KasuleTrevor Lingala Lingala 1 200 1.02 See source
KYAGABA Amharic 1 200 0.40 See source
michsethowusu Chichewa 1 200 0.09 See source
AfriSpeech-200 (Intron Health) English 1 199 0.45 CC-BY-NC-SA-4.0
Kallaama Wolof 1 197 0.20 See source
Makerere Radio Speech Luganda 1 100 0.19 See source

By subset

All 110 subsets (click to expand)
Subset Language Source Examples Hours Mean dur. (s) Max dur. (s) Dropped >30 s Dropped empty text
ach_salt Acholi Sunbird SALT 96 0.16 5.9 17.9 0 0
ach_waxal Acholi Google WAXAL 174 0.97 20.1 29.9 26 0
afr_afrikaans30s Afrikaans Afrikaans-30s 200 1.67 30.0 30.0 0 0
afr_commonvoice Afrikaans Mozilla Common Voice 129 0.23 6.3 14.0 0 0
afr_fleurs Afrikaans Google FLEURS 200 0.69 12.4 29.6 0 0
afr_makbenchmark Afrikaans ASR Africa Data Efficiency Benchmark 200 0.69 12.3 29.6 0 0
aka_waxal Akan Google WAXAL 200 1.08 19.5 28.8 0 0
amh_commonvoice Amharic Mozilla Common Voice 200 0.36 6.4 10.5 0 0
amh_kyagaba Amharic KYAGABA 200 0.40 7.2 17.0 0 0
amh_makbenchmark Amharic ASR Africa Data Efficiency Benchmark 200 0.62 11.1 18.7 0 0
amh_shunyalabs Amharic Shunya Labs 200 0.64 11.4 28.9 0 0
amh_waxal Amharic Google WAXAL 200 0.98 17.7 26.1 0 0
bam_makbenchmark Bambara ASR Africa Data Efficiency Benchmark 200 0.18 3.2 10.2 0 0
bam_robotsbambari Bambara RobotsMali Bambara ASR 200 0.19 3.4 12.2 0 0
bam_robotsmali Bambara RobotsMali Bambara ASR 200 0.12 2.1 7.9 0 0
bem_csikasote Bemba BembaSpeech 200 0.44 7.9 17.1 0 0
bem_makbenchmark Bemba ASR Africa Data Efficiency Benchmark 200 0.44 7.9 16.5 0 0
ber_tutlay Berber TutlaytAI Amazigh ASR 200 0.12 2.1 4.7 0 0
cgg_commonvoice Rukiga Mozilla Common Voice 200 0.60 10.9 27.1 0 0
dag_commonvoice Dagbani Mozilla Common Voice 200 0.26 4.6 11.4 0 0
dag_waxal Dagbani Google WAXAL 200 1.09 19.7 28.5 0 0
dga_waxal Dagaare Google WAXAL 200 1.09 19.6 30.0 0 0
eng_intronhealth English AfriSpeech-200 (Intron Health) 199 0.45 8.1 23.8 0 1
eng_salt English Sunbird SALT 96 0.13 4.9 9.4 0 0
ewe_makbenchmark Ewe ASR Africa Data Efficiency Benchmark 200 1.10 19.9 28.7 0 0
ewe_waxal Ewe Google WAXAL 200 1.07 19.3 29.8 0 0
fra_gigant French African-Accented French 200 0.13 2.4 12.0 0 0
ful_fleurs Fulani Google FLEURS 198 0.75 13.6 29.3 2 0
ful_makbenchmark Fulani ASR Africa Data Efficiency Benchmark 198 0.75 13.6 29.3 2 0
ful_waxal Fulani Google WAXAL 198 1.05 19.2 29.1 2 0
hau_africanvoices Hausa African Voices 154 0.29 6.9 11.7 46 0
hau_clearglobal Hausa CLEAR Global 200 0.40 7.3 16.6 0 0
hau_commonvoice Hausa Mozilla Common Voice 200 0.28 5.0 9.8 0 0
hau_fleurs Hausa Google FLEURS 176 0.87 17.7 30.0 24 0
hau_makbenchmark Hausa ASR Africa Data Efficiency Benchmark 200 0.17 3.0 7.4 0 0
hau_naijavoices Hausa NaijaVoices 200 0.17 3.1 11.7 0 0
ibo_africanvoices Igbo African Voices 192 0.45 8.4 16.6 8 0
ibo_commonvoice Igbo Mozilla Common Voice 89 0.14 5.6 10.3 0 0
ibo_fleurs Igbo Google FLEURS 179 0.76 15.3 30.0 21 0
ibo_makbenchmark Igbo ASR Africa Data Efficiency Benchmark 200 0.21 3.8 9.7 0 0
ibo_naijavoices Igbo NaijaVoices 200 0.23 4.1 10.7 0 0
kab_commonvoice Kabyle Mozilla Common Voice 200 0.24 4.3 10.3 0 0
kau_clearglobal Kanuri CLEAR Global 200 0.40 7.1 19.1 0 0
kik_anvkekikuyu Kikuyu African Next Voices (Kenya) 199 0.34 6.1 26.2 1 0
kin_commonvoice Kinyarwanda Mozilla Common Voice 200 0.30 5.3 10.5 0 0
kin_makbenchmark Kinyarwanda ASR Africa Data Efficiency Benchmark 200 0.31 5.6 9.9 0 0
kln_ankekalenjin Kalenjin African Next Voices (Kenya) 200 0.43 7.7 25.5 0 0
kln_commonvoice Kalenjin Mozilla Common Voice 200 0.35 6.2 15.4 0 0
koo_commonvoice Rukonjo Mozilla Common Voice 200 0.96 17.3 26.0 0 0
kpo_waxal Ikposo Google WAXAL 200 1.13 20.3 29.9 0 0
led_commonvoice Lendu Mozilla Common Voice 199 0.97 17.5 25.7 1 0
lgg_salt Lugbara Sunbird SALT 96 0.22 8.3 16.4 0 0
lin_fleurs Lingala Google FLEURS 187 0.95 18.4 29.9 13 0
lin_kasuletrev Lingala KasuleTrevor Lingala 200 1.02 18.3 26.8 0 0
lin_shunyalabs Lingala Shunya Labs 187 0.95 18.4 29.9 13 0
lin_waxal Lingala Google WAXAL 193 0.97 18.0 29.5 7 0
lth_commonvoice Thur Mozilla Common Voice 167 0.67 14.5 27.6 4 0
lug_commonvoice Luganda Mozilla Common Voice 200 0.34 6.1 10.6 0 0
lug_fleurs Luganda Google FLEURS 196 0.92 16.9 29.4 4 0
lug_makbenchmark Luganda ASR Africa Data Efficiency Benchmark 197 0.91 16.6 29.9 3 0
lug_makerereradio Luganda Makerere Radio Speech 100 0.19 6.8 15.8 0 0
lug_salt Luganda Sunbird SALT 99 0.16 5.9 11.8 0 0
lug_waxal Luganda Google WAXAL 158 0.89 20.3 29.5 42 0
luo_anvkeluo Luo African Next Voices (Kenya) 198 0.44 8.1 29.7 1 1
luo_commonvoice Luo Mozilla Common Voice 200 0.25 4.6 20.5 0 0
luo_fleurs Luo Google FLEURS 198 0.75 13.6 29.2 2 0
luy_digitaldivide Luhya Digital Divide Data 200 0.52 9.4 25.2 0 0
mlg_waxal Malagasy Google WAXAL 199 1.00 18.1 29.3 1 0
myx_waxal Lumasaba Google WAXAL 180 0.98 19.5 30.0 19 1
nbl_nextvoices Ndebele African Next Voices (Southern Africa) 144 0.64 16.1 29.4 56 0
nya_fleurs Chichewa Google FLEURS 199 0.90 16.2 29.5 1 0
nya_michsethowusu Chichewa michsethowusu 200 0.09 1.7 4.4 0 0
nyn_salt Runyankole Sunbird SALT 99 0.17 6.2 11.6 0 0
nyn_waxal Runyankole Google WAXAL 181 1.03 20.5 29.5 19 0
orm_fleurs Oromo Google FLEURS 41 0.13 11.4 25.9 0 0
orm_makbenchmark Oromo ASR Africa Data Efficiency Benchmark 200 0.41 7.4 14.2 0 0
orm_waxal Oromo Google WAXAL 198 0.88 16.0 23.9 2 0
pcm_africanvoices Nigerian Pidgin African Voices 197 0.42 7.6 19.9 3 0
pcm_commonvoice Nigerian Pidgin Mozilla Common Voice 200 0.29 5.2 9.9 0 0
ruc_commonvoice Ruruuli Mozilla Common Voice 170 0.78 16.5 29.7 30 0
rwm_commonvoice Kwamba Mozilla Common Voice 199 1.24 22.5 29.0 1 0
sna_fleurs Shona Google FLEURS 200 0.82 14.8 27.7 0 0
sna_makbenchmark Shona ASR Africa Data Efficiency Benchmark 199 1.13 20.4 29.9 1 0
sna_waxal Shona Google WAXAL 195 1.10 20.3 29.5 5 0
som_anvkesomali Somali African Next Voices (Kenya) 200 0.24 4.4 14.1 0 0
som_fleurs Somali Google FLEURS 196 0.73 13.4 26.4 4 0
som_skydheere Somali Soomali ASR 200 0.09 1.7 6.3 0 0
sot_fleurs Sotho Google FLEURS 187 0.90 17.3 29.9 13 0
sot_nextvoices Sotho African Next Voices (Southern Africa) 134 0.54 14.4 30.0 66 0
swa_commonvoice Swahili Mozilla Common Voice 200 0.33 5.9 10.2 0 0
swa_fleurs Swahili Google FLEURS 200 0.79 14.2 28.4 0 0
teo_salt Ateso Sunbird SALT 98 0.18 6.5 13.7 0 0
tsn_commonvoice Tswana Mozilla Common Voice 200 0.24 4.3 9.7 0 0
tsn_nextvoices Tswana African Next Voices (Southern Africa) 169 0.55 11.6 29.3 31 0
ttj_commonvoice Rutooro Mozilla Common Voice 199 1.02 18.4 28.5 1 0
wol_fleurs Wolof Google FLEURS 188 0.84 16.2 28.2 12 0
wol_kallama Wolof Kallaama 197 0.20 3.6 28.1 3 0
wol_makbenchmark Wolof ASR Africa Data Efficiency Benchmark 186 0.79 15.3 28.2 14 0
xho_fleurs Xhosa Google FLEURS 198 0.70 12.7 29.4 2 0
xho_makbenchmark Xhosa ASR Africa Data Efficiency Benchmark 197 0.70 12.8 26.2 3 0
xho_nextvoices Xhosa African Next Voices (Southern Africa) 132 0.62 17.0 30.0 68 0
xog_waxal Lusoga Google WAXAL 177 1.04 21.2 29.7 23 0
yor_africanvoices Yoruba African Voices 198 0.43 7.8 16.8 2 0
yor_commonvoice Yoruba Mozilla Common Voice 200 0.36 6.5 10.3 0 0
yor_fleurs Yoruba Google FLEURS 197 0.84 15.4 29.0 3 0
yor_makbenchmark Yoruba ASR Africa Data Efficiency Benchmark 200 0.20 3.6 15.5 0 0
yor_naijavoices Yoruba NaijaVoices 200 0.20 3.6 9.4 0 0
zul_fleurs Zulu Google FLEURS 188 0.82 15.6 29.5 12 0
zul_makbenchmark Zulu ASR Africa Data Efficiency Benchmark 189 0.82 15.7 29.5 11 0
zul_nextvoices Zulu African Next Voices (Southern Africa) 130 0.67 18.5 29.9 70 0

Evaluation protocol

Results reported for Sunbird/asr-whisper-51-african-languages on this benchmark use:

  • Language forcing: the decoder is given the language of each clip (language column).
  • Text normalisation: Whisper's BasicTextNormalizer (transformers.models.whisper.english_normalizer), applied to both reference and hypothesis, lower-cases and strips punctuation and symbols.
  • Metrics: corpus-level WER and CER (total edits over total reference words/characters, computed with jiwer), reported per subset, per language (pooling all of a language's subsets) and overall (pooling every clip). The pooled overall number is weighted towards languages with more clips; averaging the per-language scores instead (macro average) weights every language equally.

To reproduce the numbers exactly, apply the same normalisation and aggregate at corpus level rather than averaging per-utterance scores.

Sources and licensing

The clips are redistributed from the corpora in the source table above; the source of each clip is in its source column. Each clip remains subject to the license and terms of its upstream corpus, several of which are non-commercial (NC) or share-alike (SA). The license column shows what the upstream Hugging Face repository declares where it declares one; check the source for everything else, and before any commercial use.

For more on the corpora, see the training datasets section of the model card. The test clips here come from the held-out test splits of those same corpora.

Considerations and limitations

  • In-domain by construction. The clips come from the test splits of the same corpora whose train splits were used to train SunflowerASR, so scores reflect in-domain performance and may be optimistic for noisier or more spontaneous speech. Whether speakers are disjoint across splits depends on how each corpus was partitioned.
  • Uneven coverage. Languages range from 96 to 1,130 clips and from 1 to 6 source corpora, so per-language confidence intervals differ widely.
  • Storage-order sampling. The first 200 rows of a test split may come from a small number of speakers or sessions if the source groups rows that way.
  • Duplicated Lingala clips. lin_shunyalabs and lin_fleurs contain the same 187 recordings (identical transcripts, and waveforms that match once decoded; only the encoding differs), so Lingala's pooled metrics count them twice. Both are kept so the benchmark matches the published results; drop one of the two subsets for a de-duplicated Lingala score.
  • id is not unique. 8 ids in the Common Voice subsets for Rukiga, Rukonjo, Lendu, Thur and Ruruuli are each shared by 4 different recordings, so identify rows by position rather than by id alone.
  • Transcription conventions differ between sources (for example in casing, punctuation and how numbers are written) and are not harmonised beyond the normalisation above.

Citation

If you use this benchmark, please cite the SunflowerASR model and the upstream corpora for the languages you report.

@misc{sunbird_speech_benchmark_2026,
  title        = {Sunbird Speech Benchmark: a multilingual ASR test set for 51 African languages},
  author       = {Sunbird AI},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Sunbird/speech-benchmark}}
}
Downloads last month
10
Total file size:
2.02 GB