Apriel Voice 27B: Open-Source Model Turns Speech Directly Into Machine Commands
Apriel Voice 27B is a 27-billion-parameter open-source transformer that processes spoken requests and emits executable commands in a single pass, rather than chaining separate speech recognition and language model services. The article describes a factory engineer scenario in which a voice request to view a valve's temperature map and log a ticket above 80 °C is handled end to end, with a reported latency of about 1.2 seconds on a 4-GPU A100 node. The model combines a speech adapter over a frozen Gemma-4 backbone, allowing audio, text, and image inputs to share one attention mechanism and output function-call JSON directly. Reported results include word error rates of 6.8% on LibriSpeech and 94% correct tool-call JSON on the VocalBench suite. The model is released under an Apache 2.0 license with weights and a Docker image on the Hugging Face Hub.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in