SilencioNetwork/amharic-speech

Amharic Spontaneous Speech — Silencio

Spontaneous Amharic from 29 distinct speakers — one clip each. Four self-reported regional varieties: Addis Ababa, Gondar, Gojjam and Tigray. Long-form, mean clip length over two minutes. Audio and speaker metadata only: no transcripts, by design — human transcription is available on demand.

Hours	1.1
Clips	29
Speakers	29
Countries	2
Speaker origin regions	4
L1 speakers of the recorded language	28 of 29 (28 clips)
Audio	48 kHz stereo WAV
Mean clip length	136.1 s
Transcripts	Human-validated transcription available on demand
Licence	cc-by-nc-4.0

This release ships audio and speaker metadata only — it does not include transcripts. That is deliberate, not an omission: what it carries is speaker-level labelling on real-world spontaneous audio, which supports accent and dialect classification, speaker identification, age and gender estimation, robustness auditing and self-supervised pretraining — none of which need a transcript. Human-validated transcription with word-level alignment is available on demand over this sample or a larger subset of the same language.

Contributors answer an open prompt in their own words, on their own devices, in their own environments. Every clip is spontaneous speech, not read from a script.

Load it

from datasets import load_dataset

ds = load_dataset("SilencioNetwork/amharic-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	%
Ethiopia	28	96.6%
South Africa	1	3.4%

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

Speaker origin	Speakers	%
Ethiopia - Addis Ababa	21	72.4%
Ethiopia - Gondar	4	13.8%
Ethiopia - Gojjam	3	10.3%
Ethiopia - Tigray	1	3.4%

Demographics

Gender	Speakers	%
male	24	82.8%
female	4	13.8%
non_binary	1	3.4%

Age band	Speakers	%
25-34	16	55.2%
18-24	8	27.6%
35-44	3	10.3%
45-59	2	6.9%

Recording conditions

Device	Clips	%
Mobile	26	89.7%
Desktop	3	10.3%

Splits

Single split, test, 29 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	29 distinct
language	Language of the recording	constant: Amharic
transcript	Empty in this release — this sample ships audio and speaker metadata only	Human-validated transcription available on demand
transcript_type	Provenance of the transcript	constant: none — see transcript above
gender	Self-reported	female, male, non_binary
country	Speaker's country	Ethiopia, South Africa
mother_tongue	Speaker's self-reported first language	Amharic, Oromo, Tigrinya
dialect	Self-reported speaker origin / regional variety. This is the speaker's own background, NOT a dialect classification of the recorded language	Ethiopia - Addis Ababa, Ethiopia - Gojjam, Ethiopia - Gondar, Ethiopia - Tigray
os	Operating system of the recording device	Android, Linux, Windows, macOS
device	Recording device class	Desktop, Mobile
duration	Seconds	29 distinct
script_type	Elicitation style	constant: free_speech
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
proficiency	Speaker's self-declared proficiency in the recorded language	fluent, native

Related Amharic speech resources

Amharic has roughly 60 million speakers and is the working language of the Ethiopian federal government. Hub coverage is real but overwhelmingly read, and speaker counts are rarely stated.

Resource	Scale	Speakers	Type	Licence
google/WaxalNLP (amh_asr)	part of Waxal	not stated	Read	CC BY-SA 4.0
badrex/ethiopian-speech-flat	14.5K rows (am)	~70 across 5 languages	Read	CC BY 4.0
hadamard-2/alffa-amharic (ALFFA)	~20 h	carries speaker_id	Read	MIT
This dataset	1.10 h	29	Spontaneous, region-labelled, one clip per speaker	CC BY-NC 4.0

This does not compete on volume. What it adds is spontaneous Amharic with a regional label and a full declared language profile per speaker, in a field where the alternatives are read speech and where speaker counts usually go unreported. Mean clip length is over two minutes, so these are extended unscripted turns rather than short utterances.

Also relevant: facebook/mms-1b-all ships an amh adapter, and Whisper supports am.

Also from Silencio. Ethiopia is the origin of 5,031 hours of accented English in the corpus; a dedicated Ethiopian English release is in preparation.

What this is for

This release has no transcripts, and it is not tagged as an ASR dataset. What it carries is speaker-level labelling on real-world spontaneous audio, which supports a range of work that needs no transcript at all:

• Regional-variety classification. 4 self-reported Ethiopian regional varieties across 29 speakers, labelled per clip.
• Speaker identification and verification. 29 labelled speakers, one clip each, so any split you construct is speaker-disjoint by construction and there is no speaker leakage to control for.
• Age and gender estimation. Self-reported birth year and gender on every speaker.
• Robustness and fairness auditing. Device, OS and recording environment per clip, so performance 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.
• Not an ASR benchmark. No reference text ships with this release.

Human-validated transcription with word-level forced alignment is available on demand over this sample, over a larger subset of the same language, or as commissioned collection. The format is shipped in Kenyan Swahili, Cebuano and Tagalog.

Speaker profile

Every clip is a different contributor — one clip per speaker, no exceptions. That is the design choice this release is built around: for estimating how a model behaves across a population of speakers, the binding constraint on precision is the number of independent speakers, not the number of hours.

Accent variety is self-reported at enrolment and is not the same thing as country of origin; where a speaker reports a variety that does not match their origin, it is left as recorded rather than corrected. native_speaker and proficiency are derived from each contributor's full declared language profile rather than from a single primary-language field.

Limitations

• One clip excluded by audio QC. 30 clips were exported; 29 shipped. One was dropped for having almost no speech (5% active frames). Removed rather than substituted.

• No transcripts. Audio and speaker metadata only. This is not an ASR benchmark. Human transcription with word-level alignment is available on demand.

• Sample scale. 29 clips, 29 speakers, 1.10 hours. This is small — a probing set and a protocol demonstration, not a benchmark. Silencio holds 6,196 hours of Amharic off the shelf.

• One clip per speaker. Excellent for speaker diversity, but no within-speaker variation and no speaker-adaptive use is possible.

• Severe gender imbalance. 24 of 29 speakers are male, 4 female, 1 non-binary. Do not use this release for gender-comparative work. Balanced cohorts are available through the collection programme.

• Addis-weighted. 21 of 29 speakers report Addis Ababa; Gondar, Gojjam and Tigray have 4, 3 and 1 speakers, too few to support regional comparison.

• Accent labels are self-reported and not independently assessed.

• No acoustic annotation. Background-noise class and SNR are not annotated. Available for commissioned collection.

• 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.

• Pseudonymous speakers. speaker_id values are pseudonyms, coherent within this dataset, deliberately not linkable to speakers in other Silencio releases.

• Repeated transcript strings. 29 clips fall into 1 groups where the same transcript string appears for more than one speaker. 0 of 29 transcripts are unique. Noted so that per-clip statistics are computed with this in mind.

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, Amharic 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_amharic_2026,
  title  = {Amharic Spontaneous Speech — Silencio},
  author = {Silencio Network},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SilencioNetwork/amharic-speech}
}

Amharic off the shelf — 6,196 hours

This release is a 1.1-hour sample. The off-the-shelf Amharic inventory behind it is 6,196 hours from 7,643 contributors — already recorded, with metadata, licensable today. Direct count over the recording archive as of 9 August 2026.

Hours recorded	6,196
Recordings	409,547
Contributors	7,643

Alongside 913 hours of Tigrinya, 2,133 hours of Oromo and 5,031 hours of Ethiopian-origin English from the same contributor network.

Available by speech style, regional variety, demographic profile and recording condition. Human-validated transcription with word-level alignment is available over any subset.

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-validated transcription over a larger subset, or commissioned collection: info@silencio.network