SaarAI/asr-leaderboard-datasets

Afrivoice and Amharic configs

These configs were added by _data-prep-gsma/prep_final.py and _data-prep-gsma/prep_amharic.py in the gsma-asr-bench project. All Afrivoice rows are test-only: each config below contains exactly the held-out evaluation partition from the upstream dataset.

Common schema (identical across the three configs):

column	type	notes
file_name	string	stable per-config identifier
audio	Audio(sampling_rate=16000)	mono
duration	float64	seconds
source_lang	string (ISO 639-3)	kin, nde, or amh
target_lang	string	always eng
domain	string	acoustic / topical sub-category (see per-config notes below)
text	string	reference transcription

afrivoice_kin — Kinyarwanda (test)

• Upstream: DigitalUmuganda/Afrivoice_Kinyarwanda
• Upstream partition used: test (declared explicitly in the upstream dataset_config.yaml as {category}_test_tarred/sharded_manifests_with_image/manifest_*.json).
• Domains merged into the domain column: agriculture, education, financial, government, health.
• Language codes: ISO 639-3 kin (Kinyarwanda).
• License: CC-BY-4.0.

afrivoice_nde — Ndebele (test)

• Upstream: DigitalUmuganda/Afrivoice_V2 (Ndebele subset).
• Upstream partition used: dev_test. Note: Afrivoice_V2 does not expose a folder literally named test — the three per-domain partitions upstream are _train, _dev, _dev_test. _dev_test is DigitalUmuganda's held-out evaluation partition and is treated as the test split here (this matches the earlier prep_afrivoice_data.py in the project).
• Domains merged into the domain column: education, health.
• Language codes: ISO 639-3 nde (Northern Ndebele / Zimbabwean Ndebele).
• License: CC-BY-4.0.

amharic_amh — Amharic (acoustic-condition splits)

• Upstream: israel/g-001 (16 kHz mono Amharic ASR benchmark shipped in seven acoustic conditions).

• Upstream partition used: all 7 splits (the clean recording plus six augmentations of the same 15,965-utterance corpus).

• Splits (renamed for readability):

  upstream split (israel/g-001)	this repo (amharic_amh)	description
  train	clean_studio	Clean studio recording
  augmented	ambient_room_noise	Ambient / room noise
  over_augmented	harsh_environment	Harsh environment
  low_audio	distant_low_volume_mic	Distant / low-volume mic
  telecom	wideband_phone_volte	Wideband phone (VoLTE)
  telecom_random	randomized_phone_channel	Randomized phone channel
  over_augmented_telecom	narrowband_2g_packet_loss	Narrowband 2G + packet loss

• Language codes: ISO 639-3 amh (Amharic).

Notes on evaluation

The multilingual eval scripts in this benchmark (e.g. transformers/run_omniasr_v2_ml.sh, nemo_asr/run_eval_ml.py) default to --split test. afrivoice_kin and afrivoice_nde were originally pushed with the default train label; if you need to point the eval at them, either pass --split train explicitly, or rerun prep_final.py which now pushes them as a test-only DatasetDict.

Datasets in this leaderboard repo

Each config below corresponds to one benchmark dataset for one language. The leaderboard eval scripts in gsma-asr-bench (e.g. transformers/run_omniasr_v2_ml.sh, nemo_asr/run_eval_ml.py) select a dataset with --config_name <name> and default to --split test.

Unified row schema

All configs contributed by gsma-asr-bench follow this schema:

column	type	notes
file_name	string	stable per-config identifier
audio	Audio(sampling_rate=16000)	mono
duration	float64	seconds
source_lang	string (ISO 639-3)	e.g. kin, nde, amh, sna, zu
target_lang	string	always eng
domain	string	acoustic / topical sub-category (see per-config notes below)
text	string	reference transcription

Pre-existing SaarAI configs (e.g. waxal_sna, anv_zu, lwazi_zu) may carry additional columns from their original preparation; see the upstream sources for those.

Per-config details

afrivoice_kin — Kinyarwanda (kin)

• Upstream: DigitalUmuganda/Afrivoice_Kinyarwanda
• Upstream partition used: test
• Domains / sub-splits: agriculture, education, financial, government, health
• License: CC-BY-4.0
• Notes: Upstream dataset_config.yaml declares a test split at {category}_test_tarred/sharded_manifests_with_image/manifest_*.json — that's the exact partition used here.

afrivoice_nde — Northern / Zimbabwean Ndebele (nde)

• Upstream: DigitalUmuganda/Afrivoice_V2
• Upstream partition used: dev_test
• Domains / sub-splits: education, health
• License: CC-BY-4.0
• Notes: Afrivoice_V2 has no folder literally named test; the three per-domain partitions are _train, _dev, _dev_test. _dev_test is DigitalUmuganda's held-out evaluation partition and is treated as the test split here.

amharic_amh — Amharic (amh)

• Upstream: israel/g-001
• Upstream partition used: all 7 splits
• Domains / sub-splits: clean_studio, ambient_room_noise, harsh_environment, distant_low_volume_mic, wideband_phone_volte, randomized_phone_channel, narrowband_2g_packet_loss
• License: see upstream
• Notes: Ships in 7 acoustic-condition splits (renamed for readability). See the split rename table below.

waxal_sna — Shona (sna)

• Upstream: (SaarAI-managed; Shona speech recognition benchmark)
• Upstream partition used: test
• License: see upstream
• Notes: Shona (sna) ASR benchmark. Evaluated by transformers/run_omniasr_v2_ml.sh with --config_name waxal_sna --split test.

anv_zu — Zulu (zu)

• Upstream: (SaarAI-managed; ANV Zulu benchmark)
• Upstream partition used: test
• License: see upstream
• Notes: Zulu (zu) ASR benchmark. Evaluated by transformers/run_omniasr_v2_ml.sh with --config_name anv_zu --split test.

lwazi_zu — Zulu (zu)

• Upstream: (SaarAI-managed; Lwazi Zulu benchmark)
• Upstream partition used: test
• License: see upstream
• Notes: Zulu (zu) ASR benchmark. Evaluated by transformers/run_omniasr_v2_ml.sh and transformers/run_mmszu.sh with --config_name lwazi_zu --split test.

Amharic split rename table

For amharic_amh, the seven upstream splits were renamed for readability:

upstream split (israel/g-001)	this repo (amharic_amh)	description
train	clean_studio	Clean studio recording
augmented	ambient_room_noise	Ambient / room noise
over_augmented	harsh_environment	Harsh environment
low_audio	distant_low_volume_mic	Distant / low-volume mic
telecom	wideband_phone_volte	Wideband phone (VoLTE)
telecom_random	randomized_phone_channel	Randomized phone channel
over_augmented_telecom	narrowband_2g_packet_loss	Narrowband 2G + packet loss

Notes on evaluation

Multilingual eval scripts in gsma-asr-bench build their config name as f"{dataset}_{language}" (e.g. waxal_sna, afrivoice_kin) and pass --split test. The Afrivoice and Amharic configs added by this project follow that convention (Afrivoice) or expose named acoustic-condition splits (Amharic).