Oriloq Documentation

Fine-tuning principle

Fine-tuning consists of enriching the transcription model with annotated audio examples from your own dictations. The more representative examples of your voice and lexical field the model has, the more accurate recognition becomes for your specific terms (jargon, proper nouns, domain acronyms).

Building the dataset

Oriloq automatically saves each dictation in a configurable storage folder. The dataset is organized into two subfolders:

• •Enregistrements_a_valider — raw captures awaiting human review.
• •Enregistrements_valides — captures reviewed, corrected and marked as ready for training.

The folder path is set in Settings → Transcription → Fine-tuning dataset.

Validating recordings

The Dataset → To validate screen lets you review each recording, listen to the audio, correct the transcribed text and mark the capture as validated. Data validation is the step that brings the most accuracy gain: a model can only improve on correct examples.

Several tools speed up review: audio player with 1× to 3× speeds, integrated spell checking (since 1.0.13), automatic volume normalization, and silence compression for faster listening.

Multi-model comparison

Before validating a recording, you can compare side by side the results produced by several transcription engines (since 1.0.8). This helps identify the model best suited to your use case and re-run transcription with an alternative model if the output is unsatisfactory.

Adapter fine-tuned models

Adapter fine-tuned models (1.0.29) — the model manager shows disk size and weights size separately.

Cache and performance

Since version 1.0.13, loading a dataset entry is nearly instant thanks to the caching of the duration and hash of audio files. Validating a large dataset remains smooth even with several hundred recordings.