SilencioNetwork/amharic-speech-transcribed

Amharic Spontaneous Speech, Transcribed — Silencio

Spontaneous Amharic with human-validated transcription in Ge'ez script and word-level alignment. 45 clips from 23 distinct native speakers in Ethiopia, most of them from Addis Ababa, with a Gojjam group. Transcripts are written in Fidel (the Ethiopic script), not romanised.

Hours	0.48
Clips	45
Speakers	23
Countries	1
Speaker origin regions	3
Native Amharic speakers (declared)	23 of 23 (45 clips)
Audio	48 kHz stereo WAV
Mean clip length	38.1 s
Transcripts	human_validated: 45
Aligned words	2,958
Licence	cc-by-nc-4.0

All 45 clips carry a human-validated transcription.

Recordings are unscripted answers to open prompts, captured on contributors' own devices in their own environments. This is spontaneous speech, not read sentences.

This is a sample, not the catalogue. It is drawn to show format, audio quality and metadata, and its speaker mix does not represent what Silencio holds. The off-the-shelf catalogue behind it is far more diverse across regions, age groups, accents and recording conditions. Men are 74% of the speakers in this sample. Across the catalogue the gender split is roughly 40% female to 60% male, with men in the majority in most languages. Licensed subsets can be drawn to a specified gender, age or regional balance.

Load it

from datasets import load_dataset

ds = load_dataset("SilencioNetwork/amharic-speech-transcribed", split="test")
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	23	100.0%

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	19	82.6%
Ethiopia - Gojjam	3	13.0%
Ethiopian Amharic - Gojjam	1	4.3%

Demographics

Gender	Speakers	%
male	17	73.9%
female	6	26.1%

Age band	Speakers	%
25-34	10	43.5%
35-44	8	34.8%
18-24	5	21.7%

Recording conditions

Device	Clips	%
Mobile	44	97.8%
Desktop	1	2.2%

Splits

Single split, test, 45 rows. No train/dev/test partition is provided: at this scale a partition would leave each part too small to be meaningful.

Speakers contribute more than one clip. 12 speakers have one clip, 5 have two, 1 has three and 5 have four. Any split you build must be speaker-disjoint, which means grouping on speaker_id rather than splitting by row, or the same voice will appear on both sides.

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	23 distinct
language	Language of the recording	constant: Amharic
transcript	Human-validated transcription of the recording, Ge'ez script	45 distinct
transcript_type	Provenance of the transcript	constant: human_validated
gender	Self-reported	female, male
country	Speaker's country	constant: Ethiopia
mother_tongue	Speaker's self-reported first language	constant: Amharic
dialect	Self-reported speaker origin / regional variety. This is the speaker's own background, NOT a dialect classification of the recorded language	3 distinct
os	Operating system of the recording device	Linux, Windows
device	Recording device class	Desktop, Mobile
duration	Seconds	45 distinct
script_type	Elicitation style	constant: free_speech
words	Word-level forced alignment: text, normalised text, start and end in seconds	2,958 entries across 45 clips
n_words	Number of aligned tokens in this clip	30 distinct
age_band	Self-reported age, banded	18-24, 25-34, 35-44
native_speaker	True where the speaker declares Amharic as mother tongue or at native level in their language profile	constant: True
aligner	Method 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 Amharic	constant: 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	0.48 h	23	Spontaneous, human-validated Ge'ez transcription, word-level alignment	CC BY-NC 4.0

Use this one for evaluation, not training. It is tiny by hours. What it offers is reference text for spontaneous Amharic: the hesitations, repetitions and restarts that read-speech corpora leave out. Read-speech benchmarks overstate how well a model handles Amharic as people actually speak it.

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

Also from Silencio. Yoruba, Hausa, Kenyan Swahili, 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 native-speaker human reviewer against the audio. Recorded per clip in transcript_model.

Orthography. Transcripts are in Ge'ez script (Fidel), the way Amharic is written, not in a Latin romanisation. Sentences end with the Ethiopic full stop ።. Amharic spelling has several letters that sound the same in modern speech (for example ሀ/ሐ/ኀ, ሰ/ሠ, አ/ዐ, ጸ/ፀ), and writers do not always agree on which one to use. Normalise these homophone letters, and normalise to NFC, before computing WER or CER, or spelling differences will be counted as recognition errors.

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/amharic-speech-transcribed", 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

Every speaker gives Amharic as mother tongue and declares it at native level. Dialect is self-reported at enrolment and is left as recorded rather than corrected. Ethiopia - Gojjam and Ethiopian Amharic - Gojjam come from two versions of the enrolment form and can be merged.

Recording conditions

Contributors record on their own devices, wherever they are. Self-reported setting for the 45 clips in this release, recorded at capture:

Setting	Clips	%
Home	36	80.0%
Office	7	15.6%
Outdoor	1	2.2%
Restaurant	1	2.2%

The setting is recorded at capture but is not shipped as a column, so the schema stays identical to the other transcribed releases.

Audio quality screen. Every clip was checked for level, clipping, silence and signal-to-noise before release. The median SNR estimate is 37 dB and the lowest is 22 dB. Four further recordings were left out because they stop mid-sentence at the platform's five-minute recording limit; none were substituted. Audio is shipped at source quality and has not been resampled or denoised.

What this is useful for

• Spontaneous-speech ASR evaluation for Amharic, with human-validated reference text in Ge'ez script, in a language whose public data is overwhelmingly read speech.
• Script-aware evaluation. Native Fidel reference text for models that output Ge'ez directly rather than a romanisation.
• Forced-alignment and VAD work. Per-token timings across the whole release.

Limitations

• Small. 0.48 hours from 23 speakers is a sample for evaluation and for checking the format, not a training set.
• Mostly Addis Ababa. 19 of 23 speakers are from Addis Ababa and the rest from Gojjam. Other Amharic-speaking regions, such as Gondar, Wollo and Shewa outside the capital, are not in this sample.
• Not one clip per speaker. See Splits: split by speaker_id.
• Gender skew. 17 of 23 speakers are men, and no speaker is over 44.

Provenance and consent

Every recording is contributed by an opted-in participant through the Silencio platform, 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.

Speaker and recording identifiers in this release are pseudonymised afresh, so they do not match Silencio's internal IDs or those in any other Silencio release.

License

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

Attribution string: Silencio Network, Amharic Spontaneous Speech, Transcribed, 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: silencio.network/contact

Citation

@misc{silencio_amharic_transcribed_2026,
  title  = {Amharic Spontaneous Speech, Transcribed — Silencio},
  author = {Silencio Network},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SilencioNetwork/amharic-speech-transcribed}
}

Amharic off the shelf

This release is a 0.48-hour sample. The Amharic catalogue behind it sits in Silencio's Tier 1 production band, 10,000 hours and above. Silencio publishes catalogue depth in bands rather than exact hour counts. Bands reflect the September 2026 off-the-shelf catalogue and move as collection continues.

Tigrinya sits in the Tier 4 band, 500 to 1,500 hours, and Oromo in Tier 3, 1,500 to 5,000 hours, from the same contributor network. Ethiopia is also one of the largest single origins in the English catalogue.

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

Silencio: the catalogue behind this sample

This listing is a small fraction of what Silencio provides. Silencio specialises in underrepresented languages, accents and niche domains, delivered at scale. It draws on the largest global community of speech contributors and holds the largest off-the-shelf catalogue of niche-language speech available.

Recorded and available off the shelf. Audio already collected, with metadata, licensable today. Catalogue position as of September 2026:

Hours, off the shelf	around 500,000
Countries of contributor origin	around 180
Catalogue refresh	Weekly at record level, quarterly at catalogue level

Contributor community available on demand. Registered, consented contributors who can be activated for a specific brief:

Contributors available on demand	around 2,500,000
Countries	180+
Languages that can be collected	around 350

Anything not off the shelf can be sourced through the community. A language, an accent or dialect region, a demographic, a recording condition, a speech style or a specialist domain can be collected to a client's brief. Deep coverage across Africa, South-East Asia, South Asia and the Middle East.

Human-validated transcription is available for every language, scoped to each client's needs: script and orthography conventions, normalisation rules, word- or segment-level alignment, speaker labelling and turnaround. Every delivered clip is checked by a native-speaker reviewer. Published samples with human-validated transcripts: Kenyan Swahili, Cebuano, Tagalog / Filipino, Yoruba, Hausa and Amharic.

Proprietary and first-party. Every recording is collected directly by Silencio from consenting contributors. Nothing is scraped, and these recordings are not available in any other dataset on the internet.

Ethical sourcing, with provenance records. Every recording carries a consent record covering AI/ML training use, and contributor-level provenance documentation is available to licensees, including for EU AI Act training-data summaries. Contributors can withdraw consent, and withdrawal propagates to subsequent releases.

More at silencio.network.

For volume licensing, bespoke transcription or commissioned collection: silencio.network/contact