In our earlier post, we said the first public dataset would be Tigrinya. This is that release.
Phonetico Speech: 14.7 Hours of Tigrinya
Phonetico's first public speech dataset. CC-BY-4.0. 14.7 hours of read Tigrinya from 161 speakers, collected and verified end to end.
May 6, 2026
Phonetico Speech is a corpus of read Tigrinya speech, collected, verified, and released by Phonetico. 14.7 hours, 4,178 segments, 161 speakers. Every contributor gave informed consent and was compensated for their work. The dataset is available under CC-BY-4.0.
Every recording was made through Phonetico's collection platform by native Tigrinya speakers who knew what they were contributing to and chose to participate. Most of them live and work in Tigray.
Before contributors recorded a single segment in this release, we met with each of them over video. We explained what the data would be used for, how it would be released, and what the compensation structure was. We had them read to us live and coached them on pacing, clarity, and consistency. As recordings come in, we continue that support. We review quality, give feedback, and help contributors improve.
Evaluation splits are speaker-disjoint and gender-balanced. Twelve speakers each in dev and test, six male and six female, with no speaker overlap into training. The test split is frozen. It will not change in future versions. If you evaluate on this data and report results, they will be comparable to anyone else who does the same.
What is in the release
| Split | Duration | Segments | Speakers |
|---|---|---|---|
| Train | 11.7 h | 3,393 | 137 |
| Dev | 1.5 h | 398 | 12 |
| Test | 1.5 h | 387 | 12 |
Each segment includes the audio, transcription in Ge'ez script, duration, anonymized speaker ID, gender, word count, and speaking rate.
Versioning
Phonetico Speech uses a sequential version format. This release is v2605. Future versions will expand existing languages and add new ones. Every segment carries a version field indicating when it was introduced. Prior snapshots are accessible via git tags.
from datasets import load_dataset
ds = load_dataset("phoneticoai/phonetico-speech", "tir", split="train")
# Pin to this version
ds = load_dataset("phoneticoai/phonetico-speech", "tir", revision="v2605")
Speakers who request removal will have their segments dropped in the next version, with affected IDs documented in the changelog.
There is more to come.
Dataset: phoneticoai/phonetico-speech