Female voices
0 of 3,018 recordings
Every labelled take (1,911) is one male speaker. The 1,107 early rows with no speaker field are the same speaker.
Record some sentences →Project tracker
Roadmap
Where the dataset and the model stand today, what the next checkpoint needs, and the voices and recording conditions the dataset is still missing.
Where it stands
Counted from the public dataset on Hugging Face. A scheduled job refreshes these numbers.
3,018Recordings in the dataset
342.4Minutes of audio (305.3 train, 37.1 test)
305.3 / 360Train-split minutes toward v6
84.8% of the way, 54.7 min to go
1,429 / 2,989Lexicon terms with at least one recording
47.8% covered
5Published model versions (v1 to v5)
Figures as of 2026-10-19 12:48 UTC, dataset revision 4cb4c84b.
The next checkpoint, v6, is planned at 360 minutes of code-switched audio. The bar compares it with the dataset's train split only, since the test split is never trained on. The training job counts its minutes in the parent dataset, Read-Aloud-Voice-Sentences, so v6 may land at a slightly different point than this bar suggests.
Recordings so far
Cumulative recordings in the dataset, by the day they were recorded. Dashed lines mark the days model versions were released. Hover or tap the chart, or focus it and use the arrow keys.
Show the numbers
| Day | Recordings that day | Total | Minutes that day |
|---|
Goals for the dataset
Every recording so far is one speaker's: Daniel, an Irish-accented English speaker in Jerusalem. The model learns that voice best. These are the gaps that more recordings would close.
Different accents
1 speaker, Irish-accented English
American, Australian and South African English speakers, and any other accent: people who mix Hebrew words into their English, whether they grew up with Hebrew or learned it later.
Record some sentences →Outdoor and noisy conditions
4.4% of recordings
127 outdoors and 6 in a noisy room; the other 2,885 were recorded somewhere quiet indoors.
Record some sentences →Different microphone types
0 recordings on an external mic
All 3,018 use a built-in microphone: 2,955 on a phone (OnePlus Nord 3 5G) and 63 on one laptop (ThinkPad E15 Gen 3). Headsets, earbuds, desk and lapel mics are all missing.
Record some sentences →Work in progress
What is being built around the dataset, beyond recording more of it.
In use
The capture app
The dataset is collected with Daniel's own capture app: a phone web app in the style of Common Voice. Read a sentence, record it, review it, next. Takes are pushed to Hugging Face in batches.
Exploring
More speakers in the capture app
The app is built for one speaker today. Multi-tenancy is being explored, so that other people could record into the dataset with it too, each take tagged with its speaker.
Coming soon
Support the project
A page for donations and other ways to support the project.
Support pageModel checkpoints
Each version is the same recipe trained on more code-switched audio. v4 and v5 were published before they were benchmarked; v3 is the latest measured version. From v6 a new checkpoint is planned every 90 more minutes.
| Version | Released | Status | Code-switched min |
|---|---|---|---|
| v1 | 2026-10-15 | Benchmarked | 77.4 |
| v2 | 2026-10-15 | Benchmarked | 120.1 |
| v3 | 2026-10-18 | Benchmarked | 196.1 |
| v4 | 2026-10-18 | Published, not benchmarked | 269.3 |
| v5 | 2026-10-18 | Published, not benchmarked | 294.2 |
| v6 | Next | Planned at 360 min | 360 |
| v7 | — | Planned at 450 min | 450 |
| v8 | — | Planned at 540 min | 540 |
Minutes for v1 to v5 are the code-switched audio each version was trained on. For v6 to v8 they are the planned trigger points, not results.