SilencioNetwork/complete-voiceai-speech-dataset

Silencio Voice AI Sample Dataset

Speaker-attributed spontaneous speech. 363 labelled contributors across 111 self-reported origin varieties, recorded on their own devices in their own environments. Every clip carries origin region / variety, mother tongue, gender, device, OS, recording environment.

Hours	16.05
Clips	1,305
Speakers	363
Origin varieties	111
Languages	21
Configs	44
Speaker metadata	origin region / variety, mother tongue, gender, device, OS, recording environment
Audio	48 kHz stereo WAV, as captured
Transcripts	ASR-generated, 64% coverage, not human-reviewed — human-validated transcription available on demand
Licence	cc-by-nc-4.0

Transcripts are ASR-generated and cover 64% of clips. They have not been human-reviewed.

Human-validated transcription is available on demand for this data. Word-level forced alignment included, as shipped in cebuano-speech and tagalog-filipino-speech. Available for this sample, for a larger subset of the same language, or for commissioned collection — info@silencio.network

What this is for

Transcripts here are unreviewed ASR output — usable as pseudo-labels for semi-supervised work, but not a human-validated ASR training set. What it does carry is speaker-level labels on real-world audio, which supports a range of work needing no transcripts at all:

• Accent and dialect classification. 111 self-reported origin varieties across 363 speakers, labelled per clip. No transcripts needed.
• Spoken language identification. 21 languages in one release, recorded through the same app under the same conditions, so language is not confounded with recording setup.
• Speaker identification and verification. 363 labelled speakers. Speaker-disjoint splits can be constructed directly from speaker_id.
• Age and gender estimation. Self-reported birth year and gender on every speaker.
• Robustness and fairness auditing. Device, operating system and recording environment are recorded per clip, so error rates can be broken down by capture condition as well as by speaker attribute.
• Self-supervised pretraining. Unlabelled real-world speech is the input these methods want; the absence of transcripts is not a limitation here.

If you need transcripts, they are available on demand over this data — see below.

Load it

from datasets import load_dataset

ds = load_dataset("SilencioNetwork/complete-voiceai-speech-dataset", "amharic_ethiopia", split="train")
print(ds[0])

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

Configs: amharic_ethiopia, english_algeria, english_australia, english_belarus, english_china, english_egypt, english_french_speaking, english_haiti, english_ireland, english_jamaica, english_kenya, english_mandarin, english_medical, english_nigeria, english_pakistan, english_poland, english_russia, english_south_africa, english_uganda, english_ukraine, english_united_kingdom, english_united_states, french_canada, french_global, german_germany, global_medical, gujarati_india, hausa_nigeria, igbo_nigeria, italian_italy, japanese_japan, kannada_india, malayalam_india, mandarin_chinese_china, marathi_india, odia_india, portuguese_brazil, russian_russia, spanish_mexico, spanish_mx_mexico, telugu_india, turkish_turkey, vietnamese_vietnam, yoruba_nigeria

Speaker and recording metadata

Speaker origin / variety	Speakers	%
Nigeria - Nigerian Standard	17	9.3%
United States - Californian	16	8.8%
Canada (Québec) - Montréalais	14	7.7%
Germany - Hochdeutsch (Standard)	13	7.1%
Mainland China - Beijing	12	6.6%
unknown	11	6.0%
France - Lyonnais	10	5.5%
Nigeria - Ibadan	9	4.9%
Nigeria - Lagos Yoruba	9	4.9%
Nigeria - Pidgin-influenced	8	4.4%
Nigeria - Kano	8	4.4%
United States - General American	8	4.4%
Ethiopia - Addis Ababa	7	3.8%
France - Parisian	6	3.3%
France - Normand	6	3.3%
United States - Midwestern	6	3.3%
Mexico - Chilango (CDMX)	6	3.3%
United States - New York City	6	3.3%
Nigeria - Katsina	5	2.7%
Mainland China - Shandong	5	2.7%

Gender	Speakers	%
male	313	86.2%
female	47	12.9%
prefer_not_to_say	2	0.6%
non_binary	1	0.3%

Ethnicity (self-reported)	Speakers	%
White	134	36.9%
Black or African American	102	28.1%
Asian	70	19.3%
Other	18	5.0%
Hispanic or Latino	18	5.0%
Middle Eastern	10	2.8%
Mixed Race	7	1.9%
Native American	2	0.6%
Prefer not to say	2	0.6%

Device	Clips	%
Mobile	946	72.5%
Desktop	359	27.5%

Environment	Clips	%
home	1,031	79.0%
office	120	9.2%
outdoor	115	8.8%
public_transport	12	0.9%
cafe	9	0.7%
other	8	0.6%
library	5	0.4%
restaurant	3	0.2%
gym	1	0.1%
shopping_mall	1	0.1%

Ethnicity is self-reported by the contributor at enrolment, using a fixed category list; it is not inferred from the audio and it is not a label of the speech. It is included because accent and speaker-attribute fairness work needs it, and it is collected under the same consent as the rest of the metadata.

Human-validated transcription — available on demand

This release is not human-transcribed. Silencio provides human transcription with word-level forced alignment on demand, over this sample or over a larger subset of the same language, and as part of commissioned collection.

These releases show exactly what that deliverable looks like — human transcript text, machine forced alignment, per-token start and end times:

• SilencioNetwork/cebuano-speech — Cebuano (Bisaya) Spontaneous Speech
• SilencioNetwork/tagalog-filipino-speech — Tagalog / Filipino Spontaneous Speech

To request transcription over this data or any other Silencio subset: info@silencio.network

Limitations

• Sample scale. 1,305 clips, 363 speakers, 16.05 hours. This is a demonstration sample, not a training corpus.
• Partial, unreviewed transcripts. 64% coverage, ASR-generated.
• Gender imbalance. 313 of 363 speakers are male. Do not use this release for gender-comparative work. Balanced cohorts are available through the collection programme.
• No acoustic annotation. Recording environment, background-noise class and SNR are not annotated in this release. Available for commissioned collection.
• Audio is 48 kHz stereo WAV as captured. Resample and downmix before batching.
• 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, Silencio Voice AI Sample Dataset, 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.

Commercial licensing: info@silencio.network

Citation

@misc{silencio_complete_voiceai_speech_dataset_2026,
  title  = {Silencio Voice AI Sample Dataset},
  author = {Silencio Network},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SilencioNetwork/complete-voiceai-speech-dataset}
}

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:

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 transcription over an existing subset, or commissioned collection in a language not listed: info@silencio.network