Grăiește Moldovenește

Grăiește Moldovenește

Speech recognition and voice assistants are getting better every month. We want people in Moldova to keep up with that progress, not fall behind it.

Why

Speech recognition systems learn from recordings matched with their correct text. There are very few such recordings of Romanian as it is spoken in Moldova, with our accent, regional words and Russian loanwords. So these tools understand people in Moldova worse than other Romanian speakers. They mishear words, and voice features and services built on them work less well.

The goal of Grăiește Moldovenește is to keep up with AI development and make sure the Moldovan community doesn't fall out of the progress. Together we build a set of checked recordings and transcripts that speech recognition models can learn from and be tested on.

How it works

  1. We collect short clips of Romanian spoken with a Moldovan accent, each up to 3 seconds and a few words long.
  2. Speech recognition models suggest up to five possible transcripts for each clip.
  3. Volunteers listen to each clip and pick the transcript that matches, or type the correct one.
  4. Each clip is checked by several people independently. When most of them agree, the transcript is accepted.
  5. The accepted clips and transcripts become a dataset for training and testing speech recognition for Moldova.

What you do

  • Join with an invite link. There is no password: we email you a sign-in link.
  • Open Task and press Play (or Space). Listen as many times as you need, and slow the clip down if that helps.
  • Pick the matching transcript (keys 1–5). If none is right, choose “None of these — enter my own” (key 0) and type what you hear, with diacritics. Then press Submit (Enter).
  • Not sure? Skip the clip (S). If it's silent, not Romanian or cut off, flag it (F).
  • Each clip takes a few seconds. Do as many as you like, whenever you like. The next clip appears as soon as you submit.

Points and scoreboard

When a clip is accepted, you get 1 point if your answer matches the agreed transcript. You get 2 points if you typed the correct transcript yourself and none of the suggestions had it.

The scoreboard shows this week and all time. Only your display name is shown; your email is used only to send you sign-in links.

Sign in