This dataset is free for research, teaching and other non-commercial use under CC BY-NC 4.0. Tell us who you are and what you plan to do with it. Access is granted automatically, and we do not sell or share what you enter here.

If you want these recordings for commercial use, that is a separate licence and a conversation we welcome. Write to support@9jatesters.com or support@ranked.ng.

9jaVoice Consent-1

153 clips. 29.4 minutes, which is 0.5 hours. 13 speakers. Yoruba. FLAC, 48 kHz, mono, 16-bit. CC BY-NC 4.0.

Read-aloud speech, recorded by paid contributors on their own phones. Use it for evaluation, for fine-tuning, and as a reference set when you want to find out whether a model handles Nigerian speech at all. It is too small to pretrain on and we are not going to pretend otherwise.

Every clip here carries its own consent record. The contributor ticked an optional box, unticked by default, agreeing that the work they submitted on that task could be published in a free open dataset. Ticking it was never a condition of getting the task or getting paid. We hold a good deal more Nigerian speech than this. 153 is the number that cleared that box.

What is in it

Language	`language` value	ISO	Clips	Minutes	Speakers	Mean clip (s)
Yoruba	`yoruba`	yo	153	29.4	13	11.5
Total			153	29.4	13	11.5

The language column holds the full word, not the ISO code. ISO codes appear only in this card's metadata block, because that is what the Hub filters on.

Clips, minutes and speakers are counted from production. Mean clip length is arithmetic from the two columns beside it and nothing else. 29.4 minutes is 0.490 hours.

The per-language speaker counts add up to 13 while the release holds 13 people. Some contributors recorded in more than one language, which is ordinary in Nigeria. Speaker ids are unique across the whole release, so count languages from the table and count people from the speaker_id column.

Across 13 speakers, 29.4 minutes works out at 2.3 minutes of audio per person on average. That is enough to fine-tune or adapt a multi-speaker model. It is not enough to build a single-speaker voice.

Audio, exactly. Captured in the contributor's phone browser as WebM/Opus, which is lossy, then decoded and converted to FLAC at 48 kHz, mono, 16-bit. Nothing here is at a native rate and nothing is a lossless capture. If you run spectral analysis you will find the Opus lowpass, so plan for it rather than discovering it.

Levels: mean peak -5.6 dBFS. No clip peaks at or above -1 dBFS, measured on the source recording before conversion. These are phone recordings in ordinary rooms, not studio captures, and we publish no noise floor or SNR estimate.

Speakers. Between them they declared 9 Nigerian states, 2 genders and 3 age bands. Every contributor is paid into a Nigerian bank account matched to their own name, which is how we know who we paid. We do not verify that the person speaking on a given clip is the account holder.

How these 153 were selected

The release query is the honest part of this dataset, so here is what it checks, clause by clause.

The submission's own consent record carries a live open-release grant, joined on the grant id that submission was taken under, and not revoked. A clip can only ride the consent its own contributor gave for that task. There is no account-wide flag and no blanket terms acceptance behind this.
The submission is approved. A reviewer looked at it before it could be published. Approval checks that the submission is real work for the task. It is not a verbatim check of the audio against the prompt.
The contributor is not one whose approved written submissions span five or more job languages. See Held out, below.
A peak level is on file for the clip and it is below -1 dBFS.
The measured silence ratio is not above 0.6.

Three things that query does not require, said plainly because you will otherwise find them yourself:

It requires a duration measurement. A clip with no measured duration is excluded.
It requires a silence measurement. A clip that was never measured for silence is excluded, rather than passing by default.
It requires the work to have been PAID. Unpaid and internal qualifier work is excluded.
It requires the speaker to have separately cleared voice synthesis. That is a different permission from open release and it is checked on its own.
It requires that nobody else wrote the line. Most read-aloud lines here were built from contributors' own translations, so a clip carries two people's work: the voice and the sentence. A clip is excluded unless the person who wrote the line also gave open-release permission.
It requires a resolvable line for every read-aloud clip. A read clip whose clean line cannot be recovered is dropped rather than published with its instruction wrapper as ground truth. Spontaneous clips are kept with an empty transcript, so check speech_type or transcript != "" before training.

We have not printed per-rule exclusion counts here. If you are assessing selection bias and want them, write to us and we will send the build's own numbers.

The consent

Contributors answer two blocks before the work is submitted. A required block covering three purposes, accepted as one, and two optional boxes that stand on their own. The wording in force is version contributor-data-consent-v4.

The required block, word for word:

"Use the work I submit on this task to train and improve AI models, to build and test 9jatesters' own tools, and to include it in a dataset that 9jatesters licenses to other companies, including companies outside Nigeria. I understand 9jatesters currently stores this work on servers outside Nigeria. I can withdraw this at any time from Data permissions and 9jatesters will stop sharing my work from then on, but a company that has already trained an AI model on it cannot un-train that model."

The optional open-release box, word for word. This is the one this release selects on:

"Also publish the work I submit on this task in a free, open dataset that anyone in the world can download and keep, at no charge. I understand this one cannot be undone, because once it is published copies exist that 9jatesters cannot recall. I chose this freely and it was not required to do the task or to be paid."

The control is a checkbox that starts unticked, with its own label saying that leaving it unticked changes nothing about the task or the pay. It is not bundled into the required block, it is stored as its own record with its own scope text, and it is the only one of the permissions that is marked irreversible on our side, because publication cannot be recalled.

Two further points, because a consent claim with no way to check it is just a claim.

The permission is recorded per submission. Every row in this release carries consent_id, the id of the grant it rides on, and consent_granted_at, the moment that grant was given. Group by those columns and you can see the shape of the consent behind the set without taking our word for it.

Open publication is broader than any permission we ourselves hold. Once a clip is public under CC BY, anyone can train anything on it. That is what the wording above warned people about, in those words, before they ticked. It is also why the box was optional in the first place.

Repository layout and fields

Audio sits in one folder per language, with a single metadata.csv at the repository root joining each file to its row.

metadata.csv
data/yoruba/*.flac
data/igbo/*.flac
data/pidgin/*.flac
data/hausa/*.flac

Column	Notes
`file_name`	relative path to the clip. The filename is our internal submission id, which is what lets a clip be traced back to its own consent record
`language`	`yoruba`, `igbo`, `pidgin`, `hausa`
`transcript`	for a read clip, the exact line put in front of the speaker. Empty for every spontaneous clip, because nobody has transcribed those
`speech_type`	`read` or `spontaneous`. Filter on this before you use `transcript`
`prompt`	what was actually displayed, instruction wording and all. Context, never ground truth. For a spontaneous clip this is the question the speaker was answering
`speaker_id`	pseudonymous, stable across the whole release and across languages
`region`	Nigerian state, as the contributor declared it on their profile. Empty where the profile field was empty
`gender`	self-declared. Empty where the profile field was empty
`age_band`	self-declared band. Empty where the profile field was empty
`duration_seconds`	measured on the source recording. Can be empty
`sample_rate`	always 48000. This is the rate of the released file, written as a constant by the build, not a per-clip measurement
`peak_dbfs`	measured on the source recording before conversion, so re-measuring the FLAC will not reproduce it exactly
`rms_dbfs`	same, measured on the source
`consent_id`	id of the open-release grant this clip rides on. Opaque outside our systems
`consent_granted_at`	when that grant was given
`license`	`CC-BY-NC-4.0` on every row

No contributor's name, phone number, email address, bank details or identity documents appear in any column. Note that the scripted lines themselves are invented sentences about everyday Nigerian life, so a transcript may contain a common first name or a bank's brand name as part of the sentence. Those are written prompts, not anybody's personal data. consent_id and the submission id in file_name are internal references, published on purpose so the consent trail is checkable, and they resolve to nothing outside our systems.

We have not normalised the prompt text. Check the character inventory yourself before training, particularly Yoruba and Igbo diacritics.

How to load it

from datasets import load_dataset

ds = load_dataset("9jatesters/9javoice-yoruba", split="train")
print(ds[0]["transcript"], ds[0]["language"], ds[0]["speech_type"])

# read-aloud clips only, which are the ones with a transcript
read_only = ds.filter(lambda r: r["speech_type"] == "read")

One language:

yo = ds.filter(lambda r: r["language"] == "yoruba")

Most speech models want 16 kHz:

from datasets import Audio

ds = ds.cast_column("audio", Audio(sampling_rate=16000))

The whole set is a few hundred megabytes of FLAC. Pull it and keep it locally.

Evaluating with it

There is one split, train. We did not invent a test split, because a 13 speaker set cannot give you a speaker-disjoint one that means anything, and a random split will leak voices across the boundary and flatter your numbers.

Hold out by speaker, inside each language, so that every language has an unseen voice:

pairs = sorted(set(zip(ds["language"], ds["speaker_id"])))

held, covered = set(), set()
for lang, spk in pairs:
    if lang not in covered:
        held.add(spk)
        covered.add(lang)

test  = ds.filter(lambda r: r["speaker_id"] in held)
train = ds.filter(lambda r: r["speaker_id"] not in held)

Yoruba has 13 speaker(s), so a speaker-disjoint split removes most or all of that language. Say in your write-up which speakers you held out, otherwise nobody can reproduce your number.

One warning before you compute WER on this. transcript is the line the speaker was given, not a transcription of what they said, we have not checked every clip against its audio, and spontaneous clips have no transcript at all so they must be filtered out first. Treat any WER here as an upper bound and a smoke test rather than a leaderboard figure.

Pay and collection

Contributors join 9jatesters, verify their identity, pick up language tasks, read prompts aloud on their own phones and submit. They are paid in naira into their own Nigerian bank accounts. Cash to a bank account, not vouchers or airtime.

Items in this release were priced 25 to 150 naira each depending on the task. The rate is set by task type and never by what the contributor agreed to let us do with the work. Two people doing the same task for the same money, one ticking the open box and one leaving it, were paid identically. The one who left it blank is simply not in this file.

Across the platform, for all work and to date, contributors have withdrawn 612,440 naira to their own accounts. That is a platform-wide lifetime figure and not the cost of these 153 clips.

Held out

Contributors whose approved written submissions span five or more job languages. Their audio is held out of this release pending review by a first-language speaker of each language involved. We are not publishing a verdict on those contributors, because we have not reached one. They were paid for their work in the normal way. If a review clears them, their clips go into a later version.

Clips that clipped. A peak at or above -1 dBFS on the source recording.

Clips measured at more than 60% silence.

Anything whose contributor withdrew consent before the build.

Limits, and what sits behind them

13 speakers is a small base. Per-speaker style will correlate inside each language, so a model trained on this alone learns these voices as much as it learns the language. Behind the release sits a panel of 1,163 identity-verified contributors across 36 states, and the same pipeline produced these 153 clips, so a wider speaker base is a collection job with a number on it rather than a research problem.

Yoruba: 13 speaker(s). The thinnest slice in the set, and thin enough that any Yoruba result is indicative at best. The panel behind this release is far larger than the release, so a wider cohort in this language is a collection job rather than a research problem.

Yoruba clips average 11.5 seconds, shorter than the rest. Short utterances behave differently in ASR scoring and in TTS duration modelling. That length is a property of the prompt set.

Mostly read-aloud, with 13 spontaneous clips. 140 clips are someone reading a line they were shown, so nobody interrupts, restarts or trails off. The remaining 13 are people answering a question in their own words, and those run longer and behave like real speech. They carry no transcript. Treat the two as different material: speech_type tells you which is which.

Transcripts are the prompt line, not verbatim transcriptions. Speakers deviate, self-correct and pause, and we have not measured how often the speaker matched the line. Nothing here has been transcribed by a human listening to the audio.

There are larger open Nigerian and pan-African speech sets, including AfriSpeech-200, FLEURS and Common Voice. If what you need is volume, start there. What this set has that most of them do not is a consent record named on every row, tied to the task the submission belongs to, from a contributor who was paid in naira for the work, into a bank account verified against their own name. To be precise rather than flattering: every contributor here has a name-verified Nigerian bank account, and money accrues to their balance on approval, but our withdrawal minimum means not all of them have drawn it down yet.

If you need more speakers, more hours, more languages, verbatim human transcription, spontaneous speech, or a licensed set on terms other than open release, the panel and the pipeline that produced these clips are the ones that would produce that. Write to support@9jatesters.com or support@ranked.ng.

Personal data, and what you take on by downloading

This data is pseudonymised, not anonymous. Speaker ids are pseudonyms and we hold the mapping. Across only 13 people, region, gender and age_band together are close to unique for each speaker before anyone even hears the voice, and a voice identifies a person on its own.

Our lawful basis is the contributor's consent, given per task and recorded per submission, with the wording quoted in full above.

Our servers holding the source recordings are currently in Türkiye, not Nigeria. Contributors were told that in the required consent block, in the sentence quoted above. It is disclosed on our own trust page. Moving that storage into Nigeria is in progress and you are welcome to ask us where it stands. Publishing here is itself a transfer, made by us, to infrastructure Hugging Face operates.

CC BY-NC 4.0 grants copyright and database rights. It does not grant anything under data protection law and it cannot make your processing lawful. If you download this you are an independent controller with your own basis to establish and your own obligations, under the Nigeria Data Protection Act and under whatever law applies where you are.

Ranked Technologies Limited is registered with the Nigeria Data Protection Commission as a data controller, NDPC/DCP/14538. Contributors can ask for access, correction or objection, and can withdraw, from the Data permissions page in their own dashboard or by writing to support@9jatesters.com. They can also complain to the NDPC directly.

If a contributor withdraws, we take their clips out of this repository itself, not only out of the next version, and out of everything we license. We cannot recall a copy you have already downloaded. That is exactly what the consent wording told people before they ticked, and it is why the box was optional.

Licence

CC BY-NC 4.0. Use it for research, teaching and non-commercial work, modify it, redistribute it. Name us when you do.

We chose the NC licence deliberately, and against our own commercial interest in looking generous. The contributors who made this ticked a box saying their work could go into "a free, open dataset that anyone in the world can download and keep, at no charge". They were not asked whether a stranger could build a business on it. A Creative Commons licence cannot be revoked once someone holds a copy, so the honest move is the narrower licence, not the one that reads better on a press release.

If you want these recordings for commercial use, that is a conversation we welcome and can license separately. Write to support@9jatesters.com or support@ranked.ng.

We chose CC BY over CC0 deliberately. The contributors agreed to a free open dataset, and we want Nigerian data work to be traceable to the Nigerians who did it.

One ask that is not a licence condition, just a request from the people who recorded these clips: do not use them to build a voice that imitates a named individual, and do not use them to make anyone appear to say something they did not say.

Attribution line:

9jaVoice Consent-1, Ranked Technologies Limited (9jatesters), recorded by 13 contributors on the 9jatesters panel. CC BY-NC 4.0.

@misc{ranked2026_9javoice_yoruba,
  title        = {9jaVoice Consent-1: Consented Nigerian Speech in Yoruba, Igbo,
                  Nigerian Pidgin and Hausa},
  author       = {{ERUO FREDOLINE} and {Ranked Technologies Limited}},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/9jatesters/9javoice-yoruba},
  note         = {CC BY-NC 4.0}
}

Who made this

Ranked Technologies Limited, a Nigerian company, RC 9522220. Registered with the Nigeria Data Protection Commission as a data controller, NDPC/DCP/14538. Registered office in Lekki, Lagos. Trading as 9jatesters, the Nigerian testing and AI-data panel that collected this corpus.

Contact: support@9jatesters.com or support@ranked.ng

Founder: ERUO FREDOLINE

Write to that address to withdraw consent, to ask a licensing question, or to tell us something on this page is wrong.

The wider project

9jatesters — a panel of 1,163 identity-verified Nigerian contributors across 36 states, paid in naira into their own bank accounts for language work. This corpus came from that panel.

Ranked — our Nigerian local-business directory, and the older half of the company. It is where we learned to collect Nigerian data at scale and to verify that the people supplying it are real, which is the same problem this dataset had to solve.

NGPT — the Nigerian multilingual language model we train, and the reason this corpus exists.

9jaBench. Our published benchmark for Nigerian-language models. When our own model lost head-to-head evaluation cases against other systems, we published the losses alongside the wins.

Versions

v1.0, September 2026. First release. 153 clips, 29.4 minutes, 13 speakers, 1 languages. Every row names the open-release grant it rides on, in consent_id. A grant covers the task it was given for, so 17 grants cover these 153 clips.

Later versions will add speakers and hours as more contributors tick the open box, and may add clips currently held pending language review. Nothing can be removed from a version you have already downloaded, and we told contributors so before they agreed.

Downloads last month: 28

Total file size:

40.5 MB