Where the speech comes from
Public datasets of transcribed Darija speech are small, far too small to fine-tune a modern speech recognition model. Real Darija is also spontaneous: it mixes in French and English, and it varies from region to region. Moroccan YouTube has a great deal of exactly this kind of speech.
We built the channel list in three passes, of 89, 24, and 249 channels. Each entry recorded the channel, its topics, and a short description, and we used those notes to plan what to download. The finished downloads contain 213,089 videos under 460 channel names, about 37,000 hours of raw media. News and regional outlets account for many of the videos. Lifestyle, talk, comedy, and cooking are in the lists as well.
From raw video to clips
A downloaded video is not a training example. Intros, music, and long pauses sit between the speech, and a full episode is far longer than the 30 seconds the audio encoder reads at once. We cut the catalogue into 11,060,528 speech clips, about 33,613 hours. Early clips are windows of 30 seconds. Later clips merge short pieces of the same video so a sentence is not split across the cut. Every clip is stored as 16 kHz mono, 16-bit PCM.
That pool is larger than the set we trained on. Gemini transcribed about 10,000 hours of it. The rest stayed untranscribed, including jobs that were still queued or that failed.
What the speech is about
- Daily life20.6% · 2,079 h
- Humor7.5% · 753 h
- Culture7.0% · 707 h
- Other6.3% · 636 h
- Politics6.0% · 607 h
- Business5.2% · 530 h
- Law4.7% · 472 h
- Technology4.5% · 450 h
- Education4.3% · 438 h
- Cinema4.2% · 420 h
- Religion3.8% · 386 h
- Sports3.5% · 349 h
- Travel3.3% · 331 h
- History2.8% · 278 h
- Health2.7% · 272 h
- Music2.3% · 232 h
- Finance2.0% · 205 h
- Cuisine1.8% · 181 h
- Economy1.6% · 158 h
- Traditions1.5% · 154 h
- News1.2% · 120 h
- 11 smaller topics3.4% · 340 h
The labeling model gave each clip a topic. Hours are over the whole training corpus (about 10,099 hours). "Other" is the label for clips that fit none of the listed topics. The 11 smaller topics each account for less than 1%: automotive, science, fitness, agriculture, gaming, beauty, art, industry, engineering, aviation, nature.
Daily life is by far the largest topic, with about a fifth of the hours. Humor, culture, politics, business, law, technology, education, and cinema follow with 4% to 7% each. Specialized topics such as engineering, industry, and aviation have only a few hours each, so expect the model to struggle with their vocabulary. If you need those domains, fine-tune on additional data.
The writing convention
The labeling prompt states the language as Moroccan Darija and then gives two writing rules: write Darija words in Arabic script, and write French and English words in Latin script, as spoken (هادي هي la discipline). This isn’t the only correct way to write Darija, because no standard exists. What it gives the model is a consistent target for the two decisions that matter most: which script to use, and how to write borrowed words.
Everything else was left to Gemini: whether numbers are digits or words, whether و and ف attach to the next word, how vowels are spelled. The model picked up Gemini’s habits on all of these, and human transcribers often choose differently.
Human references, for example, often write French and English words in Arabic script (ستانفورد rather than Stanford). Our model follows our convention, so it disagrees with such references on every borrowed word. That is one reason WER is high for every system on Darija.
What we removed
Every clip from a test video
Volunteers transcribed extracts of YouTube videos, and those transcripts are our test set. We then removed from the training data every clip that came from the same videos. Removing only the transcribed extracts is not enough: other clips from the same video share the speaker, the microphone, the topic, and sometimes entire sentences. That removed 1,848 clips from 855 videos.
Labels that are too long
A 30-second clip usually has a label of a few hundred characters. Much longer labels almost always mean the labeling model got stuck repeating itself. We dropped the longest 0.5% (over 620 characters), which removed 7,441 clips.
Clips that are too short or too long
Clips under 3 seconds usually contain only a fragment of a sentence, and the audio encoder processes at most 30 seconds at a time, so the kept range is 3 to 30 seconds. Every clip that reached the final filter was already inside that range. That filter removed the 7,441 over-long labels. Together with the 1,848 clips from test videos, 1,508,646 clips remain.
Training on machine labels
Because Gemini wrote every label, the model learned Gemini’s style along with its errors. This has two consequences.
First, it probably sets a ceiling on accuracy: a model rarely learns to correct mistakes that appear consistently in its labels. Gemini is the only system that scores better than our 1.7B model on the test set.
Second, it creates a risk in evaluation: a Gemini judge might favor transcripts written in Gemini’s style. That is why our main judge comes from a different company.
What we release
The training corpus, the machine labels, and the source-video metadata stay private. The held-out test set—the 508 clips and the expert references—will be released as a public benchmark, so that other systems can be compared on the same data. The human annotation records from the judge study stay private; we report only aggregate statistics.