Kppwdfgu1/vocare-hasa

Dataset Card for Voices of Care: A Hausa Clinical Speech Benchmark

The final Hugging Face dataset URL and a citable DOI are minted at public release and are placeholders in this card until then.

Release provenance. The figures published in this document were rendered from the committed results/ mirrors at revision e684c322acb215fd35487821cbe50e9c22f561ed, released as v1.1.1; this document-level pointer supplements, and does not replace, the per-row provenance each mirror carries in its own run_id, config_hash and dataset_version columns, which remains the stronger authority for any single number. The v1.1.1 tag lands on a later commit, because a commit cannot cite its own SHA; the revision named here is an ancestor of the tagged commit.

Voices of Care is the first publicly available, clinically validated Hausa speech benchmark for community health AI. It enables rigorous, reproducible evaluation of automatic speech recognition (ASR) and intent classification on realistic Hausa clinical dialogue, so that developers, researchers, and funders can make evidence-based decisions about whether voice-based tools are ready for deployment by community health workers (CHWs) in northern Nigeria and comparable low-resource settings. Hausa is spoken by 70+ million people yet is largely absent from mainstream speech benchmarks and effectively absent from clinical speech resources; this benchmark closes that gap.

The corpus comprises 30,000 voiced instances (≈50 hours) derived from 15,000 clinically authored, PII-free sentence templates (14,966 unique Sentence_Id values in the delivered manifest) spanning ten priority health domains, recorded by ~50 native speakers stratified across four Hausa dialect clusters (Kano, Katsina, Sokoto, Zaria) with balanced gender representation. Almost every sentence was recorded twice, by two different speakers of different genders — gender-paired cross-speaker renditions, not a repeat measurement of one speaker (see Source data). It was curated by EHA Clinics Limited / eHealth Africa (Kano) with Data Science Nigeria (DSN) / EqualyzAI (Lagos) as technical partner, and funded by The Agency Fund (AI for Global Health Benchmarking Initiative).

Uses

Direct use

• ASR benchmarking and fine-tuning for low-resource Hausa clinical speech (scored with WER and CER).
• 11-class micro-intent classification of Hausa clinical utterances (scored with accuracy and macro-F1).
• Error-propagation studies for end-to-end voice pipelines (ASR hypotheses → downstream intent).
• Dialect-robustness, fairness auditing across gender/dialect, and orthography studies of Hausa hooked and glottalized consonants (ƙ, ɓ, ɗ, ƴ) and diacritics.

Out-of-scope use

The dataset must not be used to de-anonymize or re-identify speakers, to clone individuals' voices without independent consent, or as the sole basis for autonomous clinical decision-making. It is a benchmark for evaluation and research, not a validated clinical device, and it captures read speech (scripted readings), so scores are an upper bound on live spontaneous-dialogue performance.

Dataset Structure

Data instances

Each instance is a single voiced rendition of a clinically authored Hausa sentence — a native speaker reading aloud a scripted patient- or caregiver-style utterance — paired with its ground-truth Hausa transcript and de-identified metadata. Instances are not real clinical encounters.

Data fields

Every released row carries these twelve columns:

Column	Meaning
audio_id	Stable per-recording identifier; the join key to the delivered manifest.
audio	The recording itself, embedded (16 kHz mono).
transcript_raw	The manifest's Hausa transcript verbatim, unmodified.
transcript_normalized	transcript_raw under the package normalizer (see below).
normalizer_version	The normalizer that produced transcript_normalized.
sentence_id	The clinically authored sentence template this rendition reads.
intent_label	One of the 11 micro-intent classes (the intent target).
dialect	Dialect-cluster annotation (kananci, katsinanci, sakkwatanci, zazzaganci); "" on the 140 unlabelled rows below.
age_band	Speaker age band (15-29, 30-45, 45+); "" on the 140 unlabelled rows below.
gender	Speaker gender. Present on every row.
speaker_id	A deterministic pseudonym, speaker_001 … speaker_050, or "" where the speaker is unknown — see Rows with no speaker attribution.
canary	The contamination canary — see Contamination canary below.

For ASR the target is transcript_raw (scored after normalization); for intent the target is intent_label (11 classes).

Transcripts are dual and never collapsed. transcript_raw is the manifest text exactly as delivered. transcript_normalized is that text under the package's text normalizer, version 1.3.0, recorded per row in normalizer_version so a score is always attributable to the exact normalization that produced it. Both columns are always present and never null: an empty raw transcript yields an empty normalized transcript, "" on both sides. Publishing the raw column alongside the normalized one is what lets a third party re-score under their normalizer instead of inheriting ours.

Note on the audio feature and row counts. This card intentionally omits a dataset_info block and an explicit Audio feature declaration. The feature type rides in the parquet schema metadata written by voices_of_care.release.build, so load_dataset restores the Audio type from the files themselves and no in-card declaration is needed. Exact per-split byte counts are recorded at upload.

Reading the dataset without an audio codec installed

A constraint of the reader, not of this dataset, and the most likely stumble: on datasets 5.x the Audio feature decodes through torchcodec. Row indexing formats the whole row, so ds[i] decodes the audio cell and raises ImportError when that package is absent — even if you only wanted ds[i]["audio_id"]. Two access patterns work with no codec at all:

from datasets import load_dataset, Audio

ds = load_dataset("eHealthAfrica/voices-of-care-hausa", name="asr", split="test")

ds["audio_id"]                          # column access — no decode, works as-is
ds.cast_column("audio", Audio(decode=False))[0]["audio"]["bytes"]   # raw bytes

When you actually need decoded waveforms, three facts in the order most readers hit them:

• datasets 5.x with the audio codec package installed decodes every row. Install torchcodec and its FFmpeg runtime and the ordinary ds[i]["audio"]["array"] path just works. Prefer this.
• soundfile alone is not sufficient for this corpus. libsndfile offers Ogg only as Vorbis and Opus — it has no Ogg-FLAC support at any version — and a substantial minority of these recordings are Ogg-FLAC. Those cells fail with unknown error in flac decoder. Do not build a pipeline on soundfile alone; it will read most of the corpus and then stop.
• A soundfile-then-librosa chain reads everything, and is the contract this project's own evaluation campaign ran through (src/voices_of_care/asr/audio.py::_decode_audio): try soundfile first, and on failure fall back to librosa.load, which routes through audioread/FFmpeg for the formats libsndfile cannot read. Note that this audioread fallback is deprecated in librosa 0.10 and slated for removal in 1.0, so pin librosa or install the codec package for long-lived pipelines.

import io, soundfile, librosa, tempfile, pathlib

cell = ds.cast_column("audio", Audio(decode=False))[0]["audio"]
try:
    wave, sr = soundfile.read(io.BytesIO(cell["bytes"]), dtype="float32")
except Exception:
    # librosa needs a real PATH: handed a file object it uses soundfile and
    # fails identically. The suffix carries the format hint.
    suffix = pathlib.Path(cell["path"] or "cell.oga").suffix
    with tempfile.NamedTemporaryFile(suffix=suffix) as handle:
        handle.write(cell["bytes"])
        handle.flush()
        wave, sr = librosa.load(handle.name, sr=16000, mono=True)

Do not infer the codec from the file extension. .oga is roughly 90% of the corpus, far more than the share needing the fallback, because many .oga payloads are Ogg-Vorbis, which soundfile reads without complaint. Extension does not predict decodability; only trying does.

How much needs the fallback: measured over the full 68,913-cell release during the build's own verification pass, the fallback decoder was required for 1,831 cells — 2.66%, about 1 in 38. Every cell in the release was decoded, so this is a census rather than an estimate.

An earlier smoke build measured 18.8% on a 4,403-row head slice. That figure is superseded here and should not be used: the corpus is speaker-sorted, so a head slice over-represents the Ogg-FLAC subset and inflates the fallback rate roughly sevenfold. It is recorded only as a caution that a head slice is not a sample.

Data splits

The configs block above maps three subsets to Parquet globs under asr/, intent/ and v10-legacy/ (resolved at load). Counts are the committed corpus-composition record:

Config	Split	Instances
asr (default)	train	16,717
asr	validation	1,855
asr	test	5,375
v10-legacy	train	24,003
v10-legacy	validation	2,999
v10-legacy	test	2,998
intent	train	13,469
intent	validation	748
intent	test	749

Two split schemes ship, and they answer different questions:

• asr (the v1.1 primary split, default). Speaker-disjoint and text-disjoint: no speaker and no sentence template appears on both sides of a boundary, stratified by dialect and gender at speaker level under split seed 42. This is the harder and more honest protocol — a model cannot score by recognising a voice it trained on.
• v10-legacy (the frozen v1.0 protocol). Split by Sentence_Id only, so repeat renditions of a sentence never cross a boundary but speakers do. Preserved byte-reproducibly so v1.0 numbers stay checkable; not the recommended split for new work.
• intent. Split on unique Sentence_Id (0.90/0.05/0.05, seed 42) and unchanged between v1.0 and v1.1.

The intent segment sums to 14,966 — unique sentences, not recordings. It is never 30,000; that figure counts renditions, and conflating the two is the standard trap with this corpus.

Rows with no speaker attribution

140 rows carry no speaker identity, no dialect and no age band. They are 70 recordings, each present twice, whose audio_id does not appear in the upstream label release, so no demographic annotation exists for them. Their audio, transcripts, sentence_id, intent_label and gender are unaffected — gender is recoverable from the recording identifier rather than from the labels.

They are not distributed evenly across the three configs, and the difference is the point:

Config	Rows with no speaker	Why
asr (v1.1)	0	The speaker-disjoint protocol cannot place a recording with no speaker, so it excludes them by construction.
v10-legacy	140	Split by Sentence_Id only, so speaker attribution is not required.
intent	36	Split on unique Sentence_Id; 36 of the 70 survive deduplication.

This is a published property of the benchmark, not a defect discovered at release time: the campaign recorded these rows as its no_speaker_id bucket, and the corpus accounting closes exactly — 16,717 + 1,855 + 5,375 + 5,913 (dropped) + 140 (no speaker) = 30,000.

speaker_id is the empty string on these rows, and deliberately not a sentinel such as "unknown". A named value would read as a speaker who is present, and any speaker-clustered statistic would then treat 140 recordings made by unrelated people as one person's — inflating that "speaker's" weight and silently narrowing confidence intervals. The empty string is the missing form this project's own statistics code already recognises, so the failure is loud instead:

• Computing speaker-clustered statistics over v10-legacy or intent raises a deliberate ValueError on the mixed present/missing vector rather than returning a quietly wrong number. Either drop the unattributed rows explicitly, or choose utterance-level resampling explicitly.
• asr (the default, and the recommended split for new work) is unaffected — it contains none of them.

Scope of the figures above: they are exact counts over the full 30,000-recording corpus, derived from the committed corpus-composition record, not estimates or samples.

Dataset Creation

Curation rationale

The benchmark targets the intersection of an indigenous African language, a genuinely clinical register, and a downstream understanding task — a combination missing from prior resources. Ten priority clinical domains were selected against disease burden, CHW engagement, AI-deployment readiness, and scalability: maternal health, malaria, immunization, non-communicable diseases, mental health and substance abuse, pediatrics and child nutrition, infectious diseases, reproductive and sexual health, health and wellness, and environmental health.

Source data

Sentence templates were authored (clinician + AI-assisted generation) and validated for clinical accuracy and linguistic authenticity by native-Hausa speakers with community-health experience. Audio was then recorded in prompted sessions on a purpose-built platform under field-simulated conditions. Recordings were captured at 48 kHz and are resampled to 16 kHz mono for ASR.

What the two renditions are. 14,928 sentences were recorded exactly twice, and in every one of those pairs the two renditions come from two different speakers of different genders — 0 same-speaker pairs and 0 same-gender pairs in the committed pairing record. The design therefore captures cross-speaker and cross-gender pronunciation variation. It is not a test-retest repeat measurement of a single speaker, and must not be analysed as one.

Annotations

Intent labels were assigned via a two-stage process: primary annotation from the written transcript by trained Hausa-speaking annotators with community-health backgrounds, followed by independent second-reviewer adjudication on a random sample. Emotion_Tone and Speaker_Type were annotated in the same pass.

The dialect, age_band and speaker-grouping annotations are not EHA-produced: they come from the Data Science Nigeria release Data-Science-Nigeria/voice-of-care-health-dataset at the pinned immutable revision d5b34858aaa9aadf227e4671209974c0d9904b00, joined offline on a casefolded audio_id. That 40-hex revision is the exact content address the join reproduces against.

Personal and sensitive information

Sentence content is PII-free by construction (no names, dates, locations, facility names, or record numbers). Speaker identifiers were replaced with deterministic pseudonyms (speaker_001 … speaker_050) and de-linked from voice files after validation — a description of EHA's own handling, which does not make the linkage unavailable, because the upstream source publishes it; recordings were screened for incidental identifying audio and re-recorded or excluded as needed. The audio nonetheless constitutes sensitive personal data (voice/biometric) under NDPR Art. 1.3(xiii).

Pseudonymisation here is a labelling convenience, not a protection. It removes nothing: the released audio_id is the join key to the DSN source (documented above), that source is public and ungated at the revision this release cites, and its rows carry the raw speaker identifier beside the audio. This release therefore offers no speaker-identity protection beyond what the upstream source already determines, and it is not anonymous — see Limitations for both re-identification routes and what they mean for downstream use.

Consent and ethical review

Every speaker-contributor gave informed consent in Hausa with third-party verification of comprehension, documented via a plain-language Hausa information sheet, a yes/no comprehension-check consent form, a securely stored (de-linked) audio recording of the consent discussion, and a written signature or witnessed thumbprint. Contributors were told their de-identified recordings would be released openly under CC BY 4.0 for health-AI development and testing, that they could withdraw at any time without penalty, and that they would be fairly compensated. The study received formal ethical clearance from the Kano State Ministry of Health, State Health Research Ethics Committee on 13 May 2026 (clearance reference SHREC/2026/7706; committee NHREC registration NHREC/17/03//2018), with Tahir Buhari as Principal Investigator, under the National Code of Health Research Ethics (2006), NDPR (2019)/NDPA (2023), WHO guidance, and the Declaration of Helsinki (2013).

SPEAKER DEMOGRAPHIC

(Data Statements v2 section heading, used verbatim.)

Two speaker populations must not be confused, and this card keeps them apart everywhere:

• Corpus speakers — roughly 50. The full 30,000-recording corpus was recorded by approximately fifty native Hausa speakers across four dialect clusters (Kano, Katsina, Sokoto, Zaria), with balanced gender representation as a collection target.
• Evaluated speakers — nine. The v1.1 speaker-disjoint test partition contains nine speakers: six female and three male. Every headline number published for this benchmark is measured on those nine people.

This benchmark is not demographically representative, and nine speakers cannot support a representativeness claim. Per-dialect and per-gender results are reported with speaker counts and below-floor flags precisely so a reader can see how thin each cell is: the test partition's dialect cells hold 3, 3, 2 and 1 speakers respectively. Treat every disaggregated figure as an indication about these nine speakers, not an estimate for Hausa speakers, for a dialect, or for a gender.

Speaker identity is released as speaker_id, a deterministic pseudonym of the form speaker_001 … speaker_050. The index is a plain rank in sorted order over the upstream speaker-identifier set, so the mapping — although we do not publish it ourselves — is derivable by anyone holding that set, which is public and ungated. Treat speaker_id as a stable label, never as a barrier. Pseudonyms are assigned once over the whole speaker universe, so the same speaker carries the same id in every config.

Considerations for Using the Data

Bias, risks, and limitations

The full list is in Limitations below — one section, one authority. The corpus-composition risks specific to how this data was collected are:

• Modest, uneven speaker/dialect diversity. Roughly 50 speakers recorded the corpus, more from the Kano cluster than the Sokoto cluster, so per-dialect estimates are unequally reliable. (A corpus-level count; the evaluated roster is nine — see SPEAKER DEMOGRAPHIC.)
• Field-simulated, not genuine field, recording conditions.
• Class imbalance — several rare intent classes (e.g. medication concern) have single-digit test support and cannot yet be evaluated reliably; report per-class and per-subgroup results, not only aggregates.
• Deferred disaggregation — the delivered manifest carries no explicit domain column, so a per-domain × dialect × duration matrix is deferred to a future domain-tagged release.

Reference results (voc-2026-08 campaign of record)

All figures below are re-cut cell-by-cell from the audited campaign mirrors in the code repository — results/results_table.csv (ASR), results/intent_leaderboard.csv (intent) and results/anchor_calibration.csv (the anchor) — and regenerate from MLflow-tracked runs. They are measured on the v1.1 speaker-disjoint test partition of 5,375 recordings, read by nine speakers.

ASR (held-out test, 5,375 recordings; lower is better). Mean ± sample standard deviation across seeds, from results/results_table.csv:

Model	Role	Seeds	WER	CER
openai/whisper-large-v3	headline	5	15.14% ± 0.20	3.66% ± 0.05
facebook/wav2vec2-large-xlsr-53	headline	5	15.59% ± 0.12	3.60% ± 0.04
facebook/mms-1b-all	headline	5	22.81% ± 0.95	5.43% ± 0.33
facebook/mms-1b-fl102	reference (single-seed)	1	30.74%	7.12%

The two leaders (Whisper Large-v3, XLSR-53) are statistically indistinguishable. MMS-1B-fl102 carries role=reference and a single seed, so it has no dispersion and should not be read as a headline result.

On the frequently-cited 64.72% WER figure. That number is a prior-reported reference anchor for fl102 from an earlier fine-tuning run, recorded in results/anchor_calibration.csv. It is not directly comparable to anything in the table above: it was scored on a different and much smaller test denominator (1,346 examples, against this campaign's 5,375), in the fp16 era, under prior normalization choices that were not separately documented. Re-running that anchor's own configuration inside this harness reproduced 32.61 pp, a −32.11 pp difference against the recorded 64.72 — which is a measure of how far apart the two measurement setups are, not a quality verdict on either. The v1.1 campaign figure for fl102 is the 30.74% in the table. Treat every one of these gaps as a harness-comparability diagnostic.

Intent (11-class micro-intent; reference transcripts). Mean ± sample standard deviation across converged seeds, from results/intent_leaderboard.csv:

Model	Role	Seeds	Accuracy	Macro-F1
Davlan/afro-xlmr-large	headline	4 of 5	87.34% ± 1.11	68.98% ± 1.83
bert-base-multilingual-cased (mBERT)	reference	5 of 5	81.38% ± 1.17	60.71% ± 1.22

The champion's figures are a mean over the four of five seeds that converged; the fifth is excluded and the basis is stated rather than elided. The accuracy-to-macro-F1 gap reflects a sharp dependence of per-class reliability on training-set size.

Correction (2026-08-07). This card previously published the v1.0 campaign of record and its leaderboard, and the correction is recorded rather than applied silently. Corrected here: the campaign id (v1.0's voc-2026-07 → voc-2026-08); the ASR leaderboard (v1.0's 13.31 / 13.41 / 21.85 / 21.58 → the figures above — note fl102 moved upward, because the v1.1 partition is harder); the intent leaderboard (v1.0's 88.19 / 73.76 → the figures above); the ASR test denominator (v1.0's 2,998 → 5,375); the split table and split protocol (v1.0 published one Sentence_Id-split scheme, now three configs with the v1.1 speaker-disjoint primary); the speaker framing (a v1.0 corpus-level "50" now distinguished from nine evaluated speakers); and the recording design (v1.0's "two-times repeat design" was a misdescription — the pairs are cross-speaker, cross-gender renditions). The v1.0 numbers remain reproducible from the v10-legacy config.

Limitations

(Data Statements v2 maps this onto its LIMITATIONS section.)

• Nine evaluated speakers. Every headline figure is measured on nine test speakers. See SPEAKER DEMOGRAPHIC — this is a benchmark, not a population estimate.

• Read speech, not spontaneous clinical dialogue. Scores are an upper bound on live performance.

• The release is pseudonymous, not anonymous. There are two distinct re-identification routes, and the larger one is upstream of this release.

  • Primary — the public upstream join. The DSN source release publishes speaker identifiers alongside the audio, publicly and ungated, at the revision this release cites (d5b34858aaa9aadf227e4671209974c0d9904b00), and the audio_id column here is a join key to it. audio_id ships by design — the benchmark is not reproducible without it. So the speaker_NNN pseudonym is a consistency and hygiene measure, avoiding a restatement of DSN's gender- and region-encoding identifier in this release's own column; it is not a disclosure control, and we do not present it as one.
  • Secondary — small cells. dialect, age_band and gender ship as their own columns by design, because fairness auditing is impossible without them. The v1.1 test partition contains exactly one zazzaganci speaker (n = 598 recordings), so that column value alone can single out an individual for a reader with local knowledge.

  Neither route is closed by anything in this release. The no-re-identification term of the licence and the consent is what governs both, and it binds you.

• The canary detects contamination; it does not prevent it. The canary column is a detection instrument. Prior art is unambiguous on this point: BIG-bench's own canary GUID was itself absorbed into later models and is emitted by GPT-4. A canary tells you contamination happened. It cannot stop a crawler, and nothing in this release can.

• Reproducibility Limit 1 — the dependency stack is unpinned (OPEN). The declared torch and transformers floors admit a range of resolved versions, so a different resolution can move a digit. This is not fixed in v1.1; it is disclosed. Pinning is planned work, not delivered work.

• Reproducibility Limit 2 — a single observed same-artifact evaluation discrepancy (OPEN). Two evaluations of the same artifact on the same data returned +0.0405899066 pp against each other, and no framework-level cause was established. This is not fixed in v1.1; it is disclosed, and it bounds how precisely any single published digit should be read.

Both reproducibility limits remain open in this release. Neither is presented as resolved, because neither is.

Contamination canary

This release carries a contamination canary so that a curator, or a later reader of a training corpus, can determine whether Voices of Care data was ingested.

• The canary column. Every released row has a canary column. It is the empty string on every row except one.

• The GUID and the sentence. Exactly one row carries the full canary sentence, which embeds a GUID minted for this benchmark (never BIG-bench's — reusing another benchmark's GUID would make a hit unattributable):

  Canary string: the following canary GUID helps researchers detect whether Voices of Care (eHealthAfrica/voices-of-care-hausa) benchmark data was included in a training corpus: 0e993548-1ea6-4c66-92b2-e3ade1507aa1

• Which row. Stated as a rule rather than a row number, so it survives every rebuild and can be re-derived independently: min(audio_id) over the asr/train split.

Both forms ship because different readers find different things: a mechanical scan over a training corpus finds the bare GUID; a curator grepping prose finds the sentence. See Limitations for what a canary can and cannot do.

Croissant (RAI) metadata

Responsible-AI metadata using the MLCommons Croissant RAI vocabulary:

{
  "@context": {
    "rai": "http://mlcommons.org/croissant/RAI/"
  },
  "@type": "sc:Dataset",
  "rai:dataCollection": "Clinically authored, PII-free Hausa sentence templates spanning ten priority health domains were read aloud by native Hausa speakers in prompted sessions on a purpose-built recording platform under field-simulated conditions.",
  "rai:dataCollectionType": "Prompted read-speech recording of scripted clinical utterances by consented human participants; not scraped, not synthesised, and not captured from real clinical encounters.",
  "rai:dataCollectionRawData": "48 kHz audio recordings paired with the delivered manifest of transcripts, sentence ids, intent labels, emotion tone and speaker type; resampled to 16 kHz mono for release.",
  "rai:dataCollectionMissingData": "140 recordings carry no speaker identifier and are excluded from the speaker-disjoint v1.1 split; 5,913 further recordings are dropped from the v1.1 partition by the speaker-and-text disjointness constraint. Both counts are published in the corpus-composition record rather than silently absorbed.",
  "rai:dataCollectionTimeframe": "Sentence authoring, recording and annotation were carried out under ethics clearance SHREC/2026/7706, granted 13 May 2026; the benchmark evaluation campaign of record ran in 2026.",
  "rai:dataPreprocessingProtocol": "Audio resampled to 16 kHz mono. Transcripts ship in two columns: the manifest text verbatim, and a normalized form produced by the package normalizer version 1.3.0 (NFC, apostrophe canonicalisation to U+02BC, lowercasing, stray-mark removal, punctuation strip, whitespace collapse). The normalizer version is recorded per row.",
  "rai:dataAnnotationProtocol": "Intent labels were assigned in two stages: primary annotation from the written transcript, followed by independent second-reviewer adjudication on a random sample. Emotion tone and speaker type were annotated in the same pass.",
  "rai:dataAnnotationPlatform": "A purpose-built web recording and annotation platform operated by the curating institutions; dialect, age band and speaker grouping labels are sourced from the pinned upstream Data Science Nigeria release rather than annotated by the curator.",
  "rai:annotationsPerItem": "One primary intent annotation per recording, with independent second-reviewer adjudication applied to a random sample rather than to every item; inter-annotator agreement is therefore sample-based and no per-item multi-annotator distribution is released.",
  "rai:annotatorDemographics": "Trained Hausa-speaking annotators with community-health backgrounds, recruited by the curating institutions in northern Nigeria. Individual annotator demographics are not collected or released; only the professional and language profile of the annotator pool is disclosed.",
  "rai:dataUseCases": "Benchmarking and fine-tuning of automatic speech recognition and 11-class micro-intent classification for low-resource Hausa clinical speech; error-propagation studies for end-to-end voice pipelines; dialect and gender robustness auditing; Hausa orthography studies.",
  "rai:dataLimitations": "Read speech rather than spontaneous clinical dialogue, so scores are an upper bound. Only nine speakers are in the evaluated test partition. Several intent classes have single-digit test support. The dependency stack is unpinned and a single same-artifact evaluation discrepancy remains unexplained; both are open in v1.1.",
  "rai:dataBiases": "Speaker and dialect coverage is uneven, with more Kano-cluster than Sokoto-cluster speakers, and the test partition holds 3, 3, 2 and 1 speakers across its four dialect cells. Recording conditions are field-simulated rather than genuine field conditions. The benchmark is not demographically representative and must not be reported as such.",
  "rai:dataSocialImpact": "Intended to let developers, researchers and funders make evidence-based decisions about whether Hausa voice tools are ready for deployment by community health workers, and to make failure visible before deployment rather than after. Misuse as a validated clinical device, or as the sole basis for autonomous clinical decision-making, is out of scope and expressly prohibited.",
  "rai:personalSensitiveInformation": "Sentence content is PII-free by construction. Speaker identifiers are replaced with deterministic pseudonyms and de-linked from voice files. The audio nonetheless constitutes sensitive personal data (voice/biometric) under NDPR Art. 1.3(xiii), and the release is pseudonymous rather than anonymous. The speaker pseudonym is a labelling convenience, not a protection: its index is a sorted rank over a public upstream identifier set, and the released audio_id joins directly to that public source, whose rows carry the raw speaker identifier beside the audio. That upstream join is the primary re-identification route; a secondary one comes from dialect, age band and gender shipping as their own columns, which leaves small cells.",
  "rai:dataReleaseMaintenancePlan": "Versioned releases on the Hugging Face Hub with a changelog in this card and in the code repository. The test split is immutable across the v1.x line, proven by a committed assignment digest re-derived at build time. Corrections are published as new versions with an explicit correction note; released versions are not silently edited. Issues and corrections are accepted via the code repository and this dataset's Community tab."
}

This metadata is Croissant-RAI shaped: it uses the MLCommons RAI property literals and is valid JSON-LD, but it has not been validated against the mlcroissant reference implementation, which is not a dependency of this project. Read it as structured metadata, not as certified interoperability.

Additional Information

Licensing information

• Dataset: Creative Commons Attribution 4.0 International (CC BY 4.0). Use requires attribution to eHealth Africa, The Agency Fund, DSN/EqualyzAI, and participating speakers, and adherence to the consent terms (no re-identification).
• Evaluation code: Apache-2.0 (https://github.com/eHealthAfrica/voices-of-care).
• Upstream label source. The dialect, age_band and speaker-grouping labels come from Data-Science-Nigeria/voice-of-care-health-dataset at revision d5b34858aaa9aadf227e4671209974c0d9904b00. The licence variant confirmed for that release at that revision is recorded in the code repository at docs/confirmations/dsn-licence.yml, and this card's source_licence_variant metadata key is rendered from that record rather than hand-written — the card and the record cannot drift. It reads cc-by-4.0, the licence that upstream card declares at that pinned revision.

Note that the upstream source licence and this release's own CC BY 4.0 licence are two different facts that happen to name the same licence. The license field above states ours; source_licence_variant states the upstream label source's, and neither is evidence for the other.

Citation information

A citable DOI is minted at public release. Until then, please cite the dataset and the evaluation harness via the CITATION.cff in the code repository.

Contributions

Curated by EHA Clinics / eHealth Africa and Data Science Nigeria / EqualyzAI, funded by The Agency Fund. Contributions, corrections, and new baseline submissions are welcome via the code repository and this dataset's Community tab.

Changelog

Full detail lives in CHANGELOG.md in the code repository; this section records what changed on the dataset.

1.1.1

• No data changed. Identical corpus, splits, shards and row counts to 1.1.0. This release exists so the published card matches the confirmed publication records.
• The fallback-decoder share is now the corpus-wide census, 2.66% (1,831 / 68,913), replacing an 18.8% figure measured on a 4,403-row head slice of a speaker-sorted corpus. See Reading the dataset without an audio codec installed.
• source_licence_variant is confirmed as cc-by-4.0 and no longer reads TO BE CONFIRMED.

1.1.0

• Three configs instead of two. asr (default) and intent are joined by v10-legacy, which preserves the frozen, byte-reproducible v1.0 split protocol.
• A speaker-disjoint primary split. The asr config is now speaker- and text-disjoint, stratified by dialect and gender at speaker level. The v1.0 Sentence_Id-only protocol survives as v10-legacy.
• Dual transcript columns. transcript_raw (verbatim) and transcript_normalized, with the producing normalizer_version recorded per row.
• Pseudonymised speaker id. speaker_id ships as speaker_001 … speaker_050. A stable label, not a protection — the index is a sorted rank over a public upstream identifier set.
• A contamination canary column. See Contamination canary.
• Leaderboard restated. All reference results are re-cut against the v1.1 campaign of record.

1.0.0

• Initial donor-package release: two configs (asr, intent), the Sentence_Id split protocol, and the v1.0 leaderboard.