Dataset.ET Amharic Speech — v0.2.0

51.547 hours · 16,866 clips · 493 speakers · 15,443 distinct prompts

Dataset Description

• Curated by: Snapwre Technologies PLC (Addis Ababa, Ethiopia)
• Homepage: https://dataset.et
• Repository: https://github.com/snapwre/dataset-et-release
• Point of Contact: https://huggingface.co/snapwre

Dataset Summary

Read speech in Amharic, crowdsourced from volunteer contributors in Ethiopia through a Telegram bot, peer-validated by other contributors, and screened acoustically before release. Amharic has very little open speech data; this corpus exists to change that.

Contributors read a displayed prompt aloud, other contributors listen and vote on whether the recording matches, and contributors earn mobile airtime for accepted work. The corpus is built by the community it serves, and published by Snapwre Technologies PLC.

Supported Tasks

• automatic-speech-recognition — the primary use. Each clip pairs audio with the exact prompt text that was read.
• Speaker and demographic analysis, within the limits described below.

Languages

Amharic (am), written in the Ge'ez script (Ethiopic).

Dataset Structure

Data Instances

from datasets import load_dataset
ds = load_dataset("snapwre/amharic-speech", split="train")
ds[0]

{
  "audio": {"array": array([...]), "sampling_rate": 16000},
  "sentence": "…",
  "speaker_id": "spk_1a2b3c4d5e6f",
  "duration_s": 9.29,
  "speech_start_s": 0.31, "speech_end_s": 9.29,
  "gender": "male", "age_band": "18-24", "region": "addis_ababa",
  ...
}

Data Fields

Data Splits

Splits are speaker-disjoint and prompt-disjoint. No voice and no sentence appears in more than one split, so evaluation is not inflated by a model recognising a speaker or having memorised a prompt.

metadata.csv carries every field except the audio, for inspecting the corpus without downloading it.

Dataset Creation

Curation Rationale

Selected conservatively: a release that holds up matters more than a large one. Only clips with unanimous peer approval were considered, and those were then screened acoustically.

This version supersedes the previous one and contains it. 7,405 clips (22.71 h) are carried forward unchanged, keeping the same clip_id, the same speaker_id and the same split; 12,487 clips (35.0 h) are new. No speaker held out for evaluation in an earlier version has been moved into train, so scores previously reported against the earlier test split remain valid.

Source Data

Recordings are original contributions captured through the Dataset.ET Telegram bot as Opus voice messages, losslessly transcoded to 16 kHz mono FLAC. Each was made by a contributor reading a displayed prompt aloud.

Annotations

The prompt text is the transcript — contributors read a known sentence rather than transcribing free speech, so transcripts are exact by construction. Peer validators voted on whether each recording matched its prompt.

Personal and Sensitive Information

Speech is biometric. Contributors granted an explicit licence to publish their recordings in open datasets, and that consent is the basis for this release. Everything published alongside the voice has been treated as follows:

• Database identifiers never ship. speaker_id and clip_id are salted hashes, and the salt is not published.
• Demographic fields satisfy k-anonymity at k=5: every published combination of gender, age band and region describes at least five contributors. Rarer combinations were suppressed to null, so a null may mean withheld rather than not stated.
• Prompt text was scanned for phone numbers, email addresses and URLs.

Residual risk worth stating plainly: no automated check has verified that no contributor spoke identifying information aloud within a recording.

Considerations for Using the Data

Discussion of Biases

Contributors are overwhelmingly young and urban. Models trained on this will perform worse on older and rural speakers. Do not treat this as a representative sample of Amharic speakers.

Other Known Limitations

• Read speech, not conversation. Prosody, disfluency and turn-taking are unlike spontaneous speech.
• Narrow prompt domain. Prompt vocabulary skews formal and toward current affairs, so coverage of casual and conversational registers is thin.
• Telegram Opus origin. Captured through Telegram voice messages, carrying that codec's artefacts and whatever processing contributors' devices applied. Not studio audio.
• Not loudness-normalised. Normalisation is a training-time choice; the lufs column lets you do it deterministically.
• Peer validation is imperfect. During the period this data was collected, validators were paid per validation rather than per correct validation, which rewards approving quickly. Clips with any reject vote were excluded and the acoustic screening below exists to compensate, but expect residual noise.

Selection and Screening

Starting from every clip that passed peer validation, these were excluded:

Survivors were screened acoustically. 2,413 of 19,892 clips (12.1%) failed at least one check and were dropped:

A clip can fail several checks at once, so the column above sums to more than 2,413.

Additional Information

Licensing Information

Audio: CC BY 4.0. Contributors granted Dataset.ET a perpetual, irrevocable, worldwide, royalty-free licence to publish and distribute their recordings as part of open datasets, and consented to release under a permissive open data licence. Those rights are held by Snapwre Technologies PLC, which operates Dataset.ET, and are exercised here to publish the corpus openly.

Citation Information

@misc{datasetet_am_0_2_0,
  title  = {Dataset.ET Amharic Speech v0.2.0},
  author = {Dataset.ET contributors and Snapwre Technologies PLC},
  publisher = {Snapwre Technologies PLC},
  year   = {2026},
  url    = {https://huggingface.co/datasets/snapwre/amharic-speech}
}

Dataset Curators

Dataset.ET is a project of Snapwre Technologies PLC, Addis Ababa, Ethiopia. The company operates the collection platform, holds the contributor-granted rights to the recordings, and publishes the corpus openly.

Contributions

Built by 493 volunteer contributors in Ethiopia, plus the validators who reviewed their recordings. The corpus exists because they gave their voices to it.

Every shard's SHA-256, and the exact selection, screening and anonymisation policies that produced this release, are recorded in manifest.json.