Technology

Accuracy that starts at 95%+ and improves on your audio

Vatis transcribes speech it has never heard at 95%+ accuracy. What takes it past that is a training loop we run ourselves — transcribe, correct by hand, re-train, deploy — on the recordings your own team brings.

Try free — 10 minutesView pricing

Independent benchmark

Independently benchmarked against Microsoft, Google and Whisper

Researchers at the University of Bucharest scored seven transcription systems on over 126 hours of Romanian speech. Vatis finished top two of the seven overall — first on Antena 1 news and first on podcasts, and more accurate than Google on every single set.

  • ~95%

    Accurate on Antena 1 news

    95.6% on the Antena 1 news set — the highest of all seven systems tested, ahead of Microsoft, Google and every Whisper model.

  • ~90%

    Accurate on podcasts

    89.8% on podcasts — first again on unscripted conversation, and more than 12 points clearer than Google on the same audio.

  • 6 of 6

    Domains ahead of Google

    Vatis transcribed more accurately than Google's Chirp on every set in the study, from studio news to overlapping film dialogue.

Zero-shot Word Error Rate as a percentage on the in-domain and out-of-distribution test sets. Lower is better.
ModelIn-domainOut-of-distribution
ProTV — Broadcast newsAntena 1 — Broadcast newsAudiobooks — Read literary proseFilms — Overlapping film dialogueStories — Children's storiesPodcasts — Spontaneous conversation
Wav2Vec 2.027.725.640.875.454.136.3
Whisper Small31.626.840.060.041.131.9
Whisper Large12.35.914.827.3 — best in column10.9 — best in column11.7
Whisper Small + Echo9.410.118.854.121.021.6
Microsoft Transcribe2.9 — best in column4.810.6 — best in column31.117.611.5
Google Chirp (USM)12.111.320.237.622.422.4
Vatis5.24.4 — best in column13.031.216.010.2 — best in column

Figures are Word Error Rate (%) — the share of words a system gets wrong, so lower is better. It is the metric the paper reports, kept here so the table can be checked against it line for line; the headline numbers above are its flip side, accuracy. Vatis at 4.4 WER on Antena 1 is 95.6% of words correct. Every score is zero-shot: no system was tuned on these recordings before it was measured.

Outlined values are the best score in their column.

Vatis is a Romanian commercial ASR service known for its strong transcription accuracy in Romanian. Its models are trained on proprietary data and optimized for production use.
Diaconu, Vînaga & Alexe · University of Bucharest · arXiv:2603.02368
Try free — 10 minutesRead the benchmark (opens in a new tab)

Training loop

The science behind the magic

Four steps, and the fourth one leads back to the first.
  • 01

    We start with raw data

    We transcribe the audio data with the current version of our Speech-to-Text model. We split the result into fragments that can be easily analyzed, corrected, and validated. Also, we start an initial self-supervised training process for our technology at this step.

  • 02

    We do in-house labelling

    Our team of validators starts to analyse, correct and validate the data from the previous step. They take unlabelled data and label it.

  • 03

    Training, lots of it

    When we have enough new hours validated by our team, we re-train the Speech-to-Text model using a supervised technique this time.

  • 04

    Deployment

    When the training has finished, we deploy the new version of the model. We are also constantly researching better ways to improve our model's architecture.

Continuous learning

…then the cycle repeats. Constantly.

Read the case studiesSee how we compare

Every hour our validators correct becomes training data for the next version, so the model does not stand still between releases. Point the same loop at one customer's own recordings — their vocabulary, their speakers, their recording conditions — and a model tuned to them usually takes one to two months.

JURIDICE.ro

Considering that for almost 4 years we have transcribed by hand legal debates and conferences at JURIDICE.ro, Vatis Tech came as a rescue.

We like how easy the platform is to use, the fact that the transcripts are accurate and it takes a few minutes for them to be ready.

Daria NiculceaExecutive Director, JURIDICE.ro

More from Vatis

For engineers who read the docs before the marketing page

Read the documentation, try it free, tell us how it goes.

API docs (opens in a new tab)Try free — 10 minutes