SilencioNetwork/swahili-speech

Kenyan Swahili Spontaneous Speech — Silencio

Kenyan Swahili. Spontaneous speech with human-validated transcription and word-level alignment. 103 clips from 103 distinct speakers — one clip each. 99 of the 103 give Kenya as their country of origin: 66 Nairobi (Sheng-influenced), 26 Mombasa coastal, 11 diaspora and other varieties. This is not a pan-Swahili sample — Tanzania, Uganda and the DRC are effectively absent.

Hours	1.37
Clips	103
Speakers	103
Countries	2
Speaker origin regions	8
L1 speakers of the recorded language	99 of 103 (99 clips)
Audio	48 kHz stereo WAV
Mean clip length	47.8 s
Transcripts	human_validated: 103
Licence	cc-by-nc-4.0

All 103 clips carry a human transcription.

Recordings are unscripted responses to open prompts, captured on contributors' own devices in their own environments. Spontaneous speech, not read utterances.

Load it

from datasets import load_dataset

ds = load_dataset("SilencioNetwork/swahili-speech", split="train")
print(ds[0]["transcript"], ds[0]["dialect"], ds[0]["country"])

# datasets v4 returns a torchcodec AudioDecoder:
s = ds[0]["audio"].get_all_samples()
audio, sr = s.data, s.sample_rate

Requires pip install "datasets>=4.0" and FFmpeg ≥ 4.

Speaker and recording metadata

By country

Country	Speakers	%
Kenya	99	96.1%
United States	4	3.9%

Speaker origin / self-reported variety — this is the speaker's own background, not a dialect classification of the recorded language

Speaker origin	Speakers	%
Kenya - Nairobi Sheng-influenced	66	64.1%
Kenya - Mombasa	26	25.2%
United States - New York City	5	4.9%
United States - Californian	2	1.9%
Tanzania - Dar es Salaam	1	1.0%
Namibia - Namibian Afrikaans	1	1.0%
United Kingdom - RP (Received Pronunciation)	1	1.0%
United States - General American	1	1.0%

Demographics

Gender	Speakers	%
male	75	72.8%
female	28	27.2%

Age band	Speakers	%
25-34	50	48.5%
18-24	41	39.8%
35-44	11	10.7%
45-59	1	1.0%

Recording conditions

Device	Clips	%
Mobile	83	80.6%
Desktop	20	19.4%

Splits

Single split, test, 103 rows. No train/dev/test partition is provided: at this scale a partition would leave each part too small to be meaningful. Speaker identifiers are stable, so a speaker-disjoint split can be constructed at load time.

Fields

Column	Description	Values in this release
audio	Audio payload. Stored at source rate; see the spec table for the exact distribution	48 kHz stereo WAV
speaker_id	Pseudonymous speaker identifier. Coherent within this dataset; deliberately not linkable to other Silencio releases	103 distinct
language	Language of the recording	constant: Swahili
transcript	Human transcription of the recording	103 distinct
transcript_type	Provenance of the transcript	constant: human_validated
gender	Self-reported	female, male
country	Speaker's country	Kenya, United States
mother_tongue	Speaker's self-reported first language	Afrikaans, English, Swahili
dialect	Self-reported speaker origin / regional variety. This is the speaker's own background, NOT a dialect classification of the recorded language	8 distinct
os	Operating system of the recording device	Linux, Windows, macOS
device	Recording device class	Desktop, Mobile
duration	Seconds	103 distinct
script_type	Elicitation style	constant: free_speech
words	Word-level forced alignment: text, normalised text, start and end in seconds	8,391 entries across 103 clips
n_words	Number of aligned tokens in this clip	68 distinct
age_band	Self-reported age, banded	18-24, 25-34, 35-44, 45-59
native_speaker	True where mother_tongue matches the recorded language	2 distinct
aligner	Model used to produce the word timings	constant: forced-alignment
transcript_model	How the transcript text was produced	constant: human
proficiency	Speaker's self-declared proficiency in the recorded language	basic, fluent, native

Related Swahili speech resources

Swahili has roughly 200 million speakers across Kenya, Tanzania, Uganda, Rwanda, Burundi and the DRC — the most widely spoken language in sub-Saharan Africa. Hub coverage is real but thin on spontaneous speech, and thinner still on speaker count:

Resource	Scale	Speakers	Type	Licence
badrex/swahili-speech-400hr	~400 h	not stated	Aggregated, mixed provenance	CC BY 4.0
cdli/kenyan_swahili_nonstandard_speech_v1.0	32.5 h	52	Kenyan speakers with speech impairments; gated	CC BY 4.0
CLEAR-Global/Kenyan-Swahili-Speech	6 h	1	Read, Tatoeba sentences	CC BY-NC 4.0
google/fleurs (sw_ke)	3,768 utterances	few	Read Wikipedia sentences	CC BY 4.0
This dataset	1.4 h	103	Spontaneous, human-validated transcription, word-level alignment	CC BY-NC 4.0

Pick this one for evaluation, not for training. It is the smallest resource here by hours and the widest by speaker. For estimating how a model performs across a population of Kenyan Swahili speakers, the binding constraint on precision is the number of independent speakers, not the number of hours: 103 speakers contributing one clip each supports a population-level WER with a usable confidence interval and a real estimate of between-speaker variance. Six hours from a single speaker does not, at any duration.

For training volume, use the 400-hour aggregate. For a benchmark you can report a confidence interval on, use this.

Also from Silencio. Cebuano and Tagalog / Filipino are published under the same protocol.

Transcription and alignment

Two provenances, kept separate because they carry different confidence.

Text — human-validated. Every transcript was produced and checked by a human annotator against the audio. Recorded per clip in transcript_model.

Timings — machine. Word-level start and end times come from forced alignment, recorded per clip in aligner. The words column holds one entry per token with text, normalized_text, start and end in seconds:

ds = load_dataset("SilencioNetwork/swahili-speech", split="test")
for w in ds[0]["words"][:5]:
    print(f"{w['start']:6.2f}-{w['end']:6.2f}  {w['text']}")

Speaker proficiency and dialect

Proficiency comes from each contributor's own declared language profile rather than a single primary-language field. Dialect is self-reported at enrolment.

Filter on native_speaker, or on proficiency and dialect for finer control. Note the Nairobi cohort is Sheng-influenced — the urban Swahili-English contact variety — which is a distinct register from Coastal Swahili and is labelled as such.

Recording conditions

Contributors record on their own devices, wherever they are. Self-reported setting for the 103 clips in this release, recorded at capture:

Setting	Clips	%
Home	90	87.4%
Office	5	4.9%
Outdoor	5	4.9%
Gym	1	1.0%
Library	1	1.0%
Cafe	1	1.0%

The setting is captured at recording time but is not shipped as a column, to keep the schema identical to the Cebuano and Tagalog releases. It is reported here so the acoustic profile of the release is known: predominantly domestic, not studio, with a small proportion of noisier public settings. Measured SNR is not available — see Limitations.

What this is useful for

• Spontaneous-speech ASR evaluation. Unscripted Swahili with human-validated reference text, in a language whose public data is largely read or aggregated.
• Urban vs coastal Kenyan Swahili. 66 Nairobi Sheng-influenced speakers against 26 Mombasa coastal, with self-reported labels rather than inferred groupings. Sheng is the urban Swahili-English contact variety of Nairobi, and the register standard-Swahili models handle worst. As of August 2026 a Hub search for sheng returns no speech dataset named or tagged for it; we have not audited whether one exists as an unlabelled subset of a larger corpus.
• Forced-alignment and VAD work. Per-token timings across the whole release.
• Speaker-diverse benchmarking. One clip per speaker means no speaker leakage between any split you construct.

Limitations

• Sample scale. 103 clips, 103 speakers, 1.37 hours. Suitable for evaluation and alignment work; not a training corpus. Silencio holds 12,030 hours of Swahili off the shelf.
• One clip per speaker. Excellent for speaker diversity, but it means no within-speaker variation and no speaker-adaptive use.
• Gender imbalance. Skewed male; see the demographic table above. Balanced cohorts are available through the collection programme.
• Kenya-weighted, in the sample and in the archive behind it. Almost all contributors to this sample record a Kenyan origin, with a small diaspora group. The same skew holds at corpus scale: Kenya is 85.2% of recorded Swahili hours and Tanzania 12.6%, leaving Uganda, Rwanda, Burundi and the DRC barely represented despite being major Swahili-speaking populations. Treat any result from this data as Kenyan Swahili unless you have filtered otherwise.
• Proficiency and dialect are self-declared and not independently assessed.
• Word timings are machine-generated. Forced alignment, not manually corrected. Individual boundaries have not been human-verified; the transcript text has.
• Mixed audio format. Source audio is shipped untouched at its captured sample rate and channel count — see the spec table. Resample and downmix before batching.
• No diarisation. Single speaker per clip.
• Pseudonymous speakers. speaker_id values are pseudonyms, coherent within this dataset, deliberately not linkable to speakers in other Silencio releases.
• No baseline. No reference WER is published with this release.

Provenance and consent

Every recording is contributed by an opted-in participant through the Silencio app, under a consent record covering AI/ML training use. Contributors can request deletion, and deletion propagates to downstream releases. Full provenance documentation is available to licensees.

License

cc-by-nc-4.0 — free for research and non-commercial use with attribution.

Attribution string: Silencio Network, Kenyan Swahili Spontaneous Speech, 2026. CC BY-NC 4.0.

Non-commercial covers research, evaluation and publication. Benchmarking a commercial product model against this data is a commercial use and needs a licence — ask, it is usually granted for evaluation. Model weights trained on this sample inherit the non-commercial restriction. Contributors may withdraw consent; withdrawal propagates to subsequent releases but places no retroactive obligation on an existing licensee.

Commercial licensing, including terms for models trained on this data: info@silencio.network

Citation

@misc{silencio_swahili_2026,
  title  = {Kenyan Swahili Spontaneous Speech — Silencio},
  author = {Silencio Network},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SilencioNetwork/swahili-speech}
}

Swahili off the shelf — 12,030 hours

This release is a 1.4-hour sample. The off-the-shelf Swahili inventory behind it is 12,030 hours from 29,556 contributors across 56 countries of origin — already recorded, with metadata, licensable today.

Hours recorded	12,030
Recordings	1,120,060
Contributors	29,556
Countries of origin	56
Mean recording length	39 s
Mean hours per contributor	0.41

These are a direct count over the recording archive as of 9 August 2026, not an estimate or a projection. They move as collection continues, and the date is stated so any figure quoted from this card can be tied to a specific archive state.

Distribution by country of origin

Country of origin	Hours	Recordings	Contributors	Share of hours
Kenya	10,252	943,153	24,517	85.2%
Tanzania	1,515	151,107	4,359	12.6%
United States	114	10,505	299	0.9%
American Samoa	47	4,750	91	0.4%
South Africa	15	1,313	42	0.1%
Angola	12	1,164	41	0.1%
Congo	12	790	16	0.1%
Burundi	9	1,251	6	0.1%
Afghanistan	6	618	9	0.1%
United Kingdom	6	586	23	<0.1%
46 further countries	42	4,823	153	0.3%
Total	12,030	1,120,060	29,556	100%

Read the concentration honestly. The reach is 56 countries; the volume is East African. Kenya and Tanzania together are 97.8% of recorded hours, Kenya alone 85.2%. Seven countries exceed 10 hours; 33 hold less than one. If you need Tanzanian-, Ugandan- or DRC-weighted Swahili at scale, that is a commissioned collection rather than an off-the-shelf pull — the contributor network reaches those populations, the existing archive does not yet hold them in volume.

Diaspora Swahili is a real sub-holding, not noise. 236 hours from 627 contributors whose country of origin is outside East Africa, across 47 countries. Relevant if you are targeting Swahili as spoken by migrant and second-generation speakers, which differs lexically and acoustically from in-country speech.

Two caveats on the country field. It records the contributor's self-reported country of origin, not where the recording was made — someone born in Kenya and recording in London counts as Kenya. And it is unverified: entries in the long tail are consistent with occasional enrolment or device-locale errors, so treat sub-10-hour rows as indicative rather than exact. The Kenya and Tanzania figures, which carry 97.8% of the volume, are supported by dialect and language-profile data collected independently at enrolment.

Available by speech style (spontaneous, monologue, keyword), dialect region, demographic profile and recording condition. Human-validated transcription with word-level alignment — as shipped in this sample — is available over any subset.

Related East African inventory

Swahili sits inside a wider East African holding, all collected under the same consent framework and available on the same terms:

Language / variety	Hours	Contributors
Swahili	12,030	29,556
Kenyan English	5,806	12,755
Tanzanian English	1,567	1,355
Taita (Kenya)	743	541
Ugandan English	585	196
Somali	153	191
Teso	132	146
Ganda (Luganda)	76	97
Kikuyu	44	276
Luo	17	80
Meru	9	36
Kamba	10	62
Luyia	6	34
Kalenjin	5	46

Several of these — Taita, Teso, Kikuyu, Luo, Kamba, Luyia, Kalenjin — have no dedicated speech dataset on the Hub at all. They are collected to order through the same contributor network.

Silencio corpus and collection network

Two distinct figures, because they answer different questions.

Recorded and available off the shelf — audio already collected, with metadata, licensable today. Swahili alone accounts for 12,030 hours from 29,556 contributors:

Hours recorded	127,793
Recordings	9,392,870
Contributors who recorded	222,145
Languages	156
Countries and territories of origin	216

Contributor network available for commissioned collection — registered, consented contributors who can be activated for a specific brief. These are not active contributors to the corpus above; they are the pool it is drawn from and extended through:

Registered contributors	2,000,000+
Countries	180+
Languages reachable	250+

For volume licensing, human-validated transcription over a larger Swahili subset, or commissioned collection in a language not listed: info@silencio.network