Products / OreaCuraçao
A radio host in a small Curaçao studio speaking into the microphone during a live talk show.
Speech to text · Papiamentu

Orea

Speak Papiamentu and read it back as text, including the Dutch and English that slip in along the way.

Galvan IntelligenceShipped · V1.12

Orea Studio includes 120 free minutes a month. Or speak into it in your browser at labs.galvangroup.co/orea

  • 01Built for switching between languages
  • 02Speakers separated, every line timed
  • 03Doubtful words marked, with alternatives
  • 04Apache 2.0, weights and base alike

Speech recognition for how Curaçao talks.

Orea turns spoken Papiamentu into text.

People on Curaçao switch between Papiamentu, Dutch and English in the middle of a sentence. That switching is the hardest part for a speech model, and it is what Orea was trained for.

It is an open speech model, trained further on Papiamentu. The browser demo runs the live model.

For longer work there is Orea Studio. Upload a meeting or an interview. You get back who said what and when. Words it is unsure of are marked, so you check those first.

orea.galvangroup.co · Transcript
Orea Studio's transcript view: a waveform with playback speed, then each line timed and written as spoken, with unsure words underlined and a guide to the marks on the right.

Orea Studio: a timed transcript, with the words Orea is unsure of underlined.

  • In the browserSpeak into it

    Clips of up to 60 seconds, transcribed while you wait. Mark a result right or wrong, and send a correction.

  • Orea StudioLong recordings

    Upload a file, name the speakers, edit in place, and export as SRT, VTT, text, JSON or Word, or share by link.

  • Inside PreceptSpoken questions

    Precept, our law research portal, takes spoken questions through Orea.

Orea at work.

A real screen from Orea Studio, captured on 23 October 2026: Somesh's language test, transcribed by Orea V1.12.

From voice to transcript, in five steps.

What happens between pressing record and holding a transcript you can use.

Step 01 of 05

You speak, or upload

In the browser, up to 60 seconds, or a file in Studio.

So Orea knows what it is listening to before it starts.

Step 02 of 05

It recognises the speech

The fine-tuned model writes what it hears, switching included.

So You read what was said, not a translation of it.

Step 03 of 05

It separates speakers and times every line

Studio separates the voices and times every line.

So You find who said what, and when, without listening to the whole recording again.

Step 04 of 05

It marks the doubtful words

Words it is unsure of are flagged, with alternatives.

So You check the marked words first, instead of reading every line with the same suspicion.

Step 05 of 05

You edit, export and improve it

SRT, VTT, text, JSON or Word, or share by link.

So Subtitles, minutes and archives come out of the same transcript, and your corrections can improve the next release.

Licence

Open licence, so it can run on your own machines.

The weights and the base model are both under Apache 2.0, which is what lets us run Orea on a setup reserved for you, or on your own machines.

Every release, measured the same way.

Word error rate: the share of words that come out wrong. Lower is better.

  1. V1
  2. V1.1
  3. V1.2
  4. V1.4
  5. V1.5
  6. V1.6
  7. V1.7
  8. V1.9
  9. V1.10
  10. V1.11
  11. V1.12

Weighted word error rate per release. V1.2 was withdrawn after four hours because it sometimes translated Papiamentu into English. V1.3 and V1.8 were never served.

Each release is measured on the same test set, and the error rate has fallen with every one.

Prepared speech comes out cleanest. Spontaneous speech and meetings are harder.

The test set favours the way people actually talk over read-aloud speech.

Last measured on 2026-09-30.

Full measurements at Galvan Intelligence →

For anything said in Papiamentu.

Uses that Orea and Orea Studio support today.

Meetings

Minutes from a recording

Upload the recording and get the speakers named and every line timed, ready to turn into minutes.

Research

Interviews that switch

Transcribe interviews that move between Papiamentu, Dutch and English without losing the Dutch phrases.

Video

Papiamentu subtitles

Export SRT, VTT, text, JSON or Word, and put Papiamentu subtitles on a video.

Your systems

Voice input

Precept already takes spoken questions through Orea. The same model can take voice input in a system built for you.

Where we draw the line.

Where it is strong, where it is not, and what happens to what you upload.

  • It is a Papiamentu model.

    Dutch and English inside Papiamentu speech are measured. Other languages are opened as they are measured.

  • Spontaneous speech is harder.

    Prepared speech comes out cleaner than meetings. Check the marked words before you publish a transcript.

  • Browser clips are short.

    In the browser, a clip can be up to 60 seconds, with eight recordings a minute of up to 2 MB each. Longer recordings go to Orea Studio.

  • Your recordings may train the next release.

    What you upload may help train the next Orea release. Only upload recordings you have the right to process.

Shipped, and still improving.

V1.12 is the live release. Earlier releases are listed on Galvan Intelligence.

  • Orea V1.12The live release, fine-tuned for PapiamentuShipped
  • Orea StudioLong recordings, speakers, timing, export. 120 free minutes a monthShipped
  • Browser demoClips up to 60 seconds at labs.galvangroup.co/oreaShipped
  • Other languagesOpened as they are measuredAs measured

Use it today.

Open Orea Studio at orea.galvangroup.co, or speak into it at labs.galvangroup.co/orea. For a setup reserved for you, write to us.

Orea

Hear your recordings as text.

Start in Orea Studio with 120 free minutes a month, or write to us about running Orea for your organisation.

Write directly [email protected]
Open Orea Studio
See where this fits in your business: book an assessment Papiamentu, with the Dutch and English left in.