Transcription for language research and corpora

Create first-pass transcripts for fieldwork, interviews and corpora in 100 languages, with word-level timestamps and speaker turns in JSON that you can convert for your annotation tools.

Get an API key Quickstart Supported languages

Last updated: · Markdown

Why researchers use it

  • 100 language codes, including lower-resource languages such as Yiddish (yi), Amharic (am), Swahili (sw), Yoruba (yo), Māori (mi) and Hawaiian (haw). See the full list.
  • Word-level start and end times for alignment work.
  • Speaker labels (include_diarization) for interview turns, or physical-channel labels for multi-microphone recordings.
  • Translation of the transcript into a working language on the Oracle models, for teams that do not all read the source language.

A request for an interview recording

{
  "audio_url": "https://example.org/corpus/session-014.wav",
  "model": "oracle",
  "language": "yi",
  "word_timestamps": true,
  "include_diarization": true,
  "num_speakers": 2,
  "output_format": "json"
}

The JSON result gives segments[] with start, end, speaker, text and words[]. Converting that to ELAN, Praat TextGrid or CSV is a short script on your side; the API does not export those formats directly.

Check the output before you analyze it

  • Machine transcripts are a first pass. Plan human correction for anything you publish or annotate.
  • Accuracy varies by language, dialect, recording quality and domain, and is lower for languages with little training audio. Measure word error rate on a sample from your own corpus before you transcribe all of it.
  • Set language explicitly for code-switched recordings; auto-detection picks one language per file.
  • Normalize orthography (diacritics, spelling conventions) on both sides before computing error rates.

Budget

A 100-hour corpus costs US$48 on base or US$144 on oracle, metered per second. Research groups that need invoicing or volume terms can contact KolWrite Enterprise.

Questions and answers

Does KolWrite export Praat or ELAN files?

No. It returns JSON with segment and word times, which converts easily to TextGrid, EAF or CSV.

Can I trust transcripts of lower-resource languages?

Treat them as drafts. Test a sample, measure error rates, and plan review.

Is my research audio used to train models?

KolWrite states on its models page that audio submitted through the API is not used to train its models.

Related pages