Introduction
This transparency note explains how the Acai voice agent works in Oral Practice and Oral Reflection. The voice agent is an AI conversation partner that students can speak with to practice or reflect out loud. It is an optional feature: teachers choose whether to use a voice-agent mode, and we recommend offering students an alternative where possible.
The voice agent is never part of an Oral Assessment: live assessments are always led by human assessors.
Functionality
The voice agent holds a spoken conversation with the student. It asks questions and follow-up questions based on the assignment description, the configured topics, and — when a submission step is included — the student’s own submitted work, which allows it to personalize the conversation.
Three AI services work together:
- Reasoning — the conversation is driven by Claude Opus 5, accessed through Amazon Bedrock with regional hosting.
- Speech synthesis — the agent’s voice is generated by Cartesia.
- Transcription — speech is transcribed as described in the Transparency Note: Transcription in Oral Assessment & Practice.
Conversation inputs processed for reasoning are handled entirely within AWS infrastructure and are not sent to Anthropic. Beyond limited automated abuse detection performed by AWS, they are not retained after processing, and they are never used to train, retrain, or fine-tune any model.
The voice agent speaks English. It understands English, Spanish, French, Portuguese, German, Italian, and Dutch.
User Impact
Students are informed through the user interface, during onboarding and before their first recording, that they will be speaking with an AI agent. A persistent visual cue keeps this clear throughout the session. The agent’s questions focus on the student’s learning process — how they approached their work and how they reason about it — with follow-up questions that give students room to reformulate and clarify.
Pedagogical Foundation
The voice agent’s conversational approach draws on established educational research, including Bloom’s taxonomy, the SOLO taxonomy, research on metacognition and reflective practice, and guidance on interactive oral assessment. For the full research foundation and sources behind the voice agent’s conversational approach, see The Research Behind the Acai Voice Agent.
Limitations and Risks
- The agent may take a few seconds to respond, particularly at the start of a conversation.
- It can misunderstand what a student says, especially with poor audio or heavy background noise.
- The agent’s questions are designed to help students articulate their learning; they are never judgments about the student or the quality of their work.
- It speaks English only, which may disadvantage students in programmes taught in other languages; teachers should weigh this when choosing a voice-agent mode.
Transparency, Privacy, and Data Governance
Recordings and transcripts of voice-agent conversations are stored on Microsoft Azure and follow our standard retention policy. In Oral Practice they remain private to the student; in Oral Reflection they are accessible to the teacher. Students can download and delete their own recordings.
Human Agency and Oversight
The voice agent never grades and never judges. In Oral Practice, conversations are fully formative, with no grades and no teacher reporting. In Oral Reflection, any review, feedback, or grading is done by the teacher.