Where the speech comes from, how raw video became clips, the cleaning steps, and what it means to train on machine labels.
Read the pageTechnical details
The full technical story.
The product blog tells the story in one pass. These 12 pages expand each part of it, in the same order. The indented pages go one level further into the method.
How the model is built, how we made the Darija vocabulary, the three training stages, and the control models that isolate the speedup.
Read the pageSix kinds of spelling difference that WER counts as errors, how large the effect is, and what text normalization can and can't fix.
Read the pageThe study with seven native speakers, the SER formula, the three tests, the results for each system, and the cases where the judge is wrong.
Goes further
All nine systems on the 508 test clips, under each judge and under WER and CER, with the comparison to MoulSot and Voxtral and the speed measurements.
Goes further
The two test sets, the scoring protocol, Voxtral's published figures next to our own rerun, and character, word, and meaning error for every system.
Read the pageInstalling llama.cpp, downloading the model, transcribing files, and where the laptop speed comes from.
Read the page