Voice assistants today often use separate systems for listening, understanding, and talking, which can cause delays and errors. Apriel Voice is a single model that directly processes spoken requests, understands context, holds conversations, and can take actions like opening a ticket or calling a tool. It is built from an existing text-and-image model and trained in two stages to add speech abilities while keeping its other skills.
The researchers trained Apriel Voice in two stages. First, they taught the model to understand speech by practicing translation and transcription. Then, they fine-tuned the whole model on many tasks including spoken dialogues and tool use, using a frozen copy of the original model to prevent forgetting.
In tests, Apriel Voice scored 98.90 on Big Bench Audio for spoken reasoning and completed 65.51% of tasks in a voice test. This matters because it shows a single open model can handle real-world voice interactions without needing separate pieces.