AI models

Different models prepare transcripts, comments, rubric assessments, camera observations, and live speech.

Models by step

The Studio uses several AI models, each for one step, and tries them in a fixed order when one fails. The student's browser records the picture and sound. Transcription produces the words and times used for written feedback.

See the models for transcription, written feedback, rubric grading, camera observations, and live voices.

Transcription of an ordinary recording

First modelElevenLabs Scribe, scribe_v1
Next modelOpenAI whisper-1, then AssemblyAI
What it readsRecording file with picture and sound
Select a step to see its models and the material they receive.

Written feedback settings

AI feedback model appears under Advanced in Course settings. Best available model is recommended and starts with Luna. The other option starts with gpt-oss-120b.

The Studio requests Groq gpt-oss-120b after a failed Luna request. Cloudflare Workers AI gpt-oss-120b follows as needed.

The same models support the task-fit check, the comment composer, and Draft from my topic. They also answer questions under Ask about your students' videos on the Grades page's Class trends tab. The Studio checks quoted words before accepting a rubric score. You can review each line's reason and recorded moments.

Other analysis and room video

Delivery analysisNVIDIA Nemotron 3 Nano Omni assesses delivery when the course uses it.
Tone shiftsDistilBERT, distilbert-sst-2, on Cloudflare Workers AI classifies the sentiment of transcript words.
Words on screenTesseract reads text shown on the recorded screen.
Speaking measuresThe Studio calculates pace, pauses, and filler counts from the transcript and audio.
Live room videoCloudflare RealtimeKit transmits the meeting and creates its shared recording.