Introduction
This transparency note explains how automatic transcription works in the Oral Assessment & Practice tools: Oral Practice, Oral Assessment, and Oral Reflection. It applies to all three tools.
Every recording made in these tools is transcribed automatically. Transcription starts together with the recording and cannot be disabled independently.
Functionality
Transcription converts the spoken conversation into text in real time. The transcript powers several features:
- a full, clickable transcript alongside the recording — clicking a word jumps to that moment,
- keyword search within a session, and
- Acai session highlights, which link to notable moments in the recording.
Transcription is provided by Mistral’s Voxtral Mini Transcribe Realtime model, hosted in the EU with zero data retention: audio is processed to produce the transcript and is not stored by the provider, and is never used to train, retrain, or fine-tune any model.
Transcription supports English, Spanish, French, Portuguese, German, Italian, and Dutch.
User Impact
Who can see a transcript follows who can see the recording. In Oral Practice, recordings and transcripts are private to the student. In Oral Assessment and Oral Reflection, teachers and assessors can access recordings and transcripts as part of reviewing and giving feedback.
Limitations and Risks
Automatic transcription is not perfect. Accuracy can be affected by audio quality, background noise, accents, and specialized terminology. When multiple people share one microphone — for example during an in-room oral assessment — the transcript may not attribute speakers correctly. Transcripts should be treated as an aid to reviewing the recording, not as a verbatim legal record.
Correction Options
Transcripts cannot be edited, and cannot be deleted separately from the recording. Deleting a recording deletes its transcript. Where a transcript is inaccurate, the recording itself remains the authoritative source.
Transparency, Privacy, and Data Governance
Recordings and transcripts are stored on Microsoft Azure, like other user-generated content in FeedbackFruits, and follow our standard retention policy. FeedbackFruits claims no ownership of recordings; students can download and delete their own recordings. No conversation data is used to train AI models.
Human Agency and Oversight
Transcripts, and the features built on them such as keyword search and Acai session highlights, support human review. They never produce grades or judgments: assessment decisions are always made by teachers and assessors. Students and teachers are informed in advance whenever an AI agent takes part in a conversation, and the design and framing of the interaction make its AI nature evident throughout.