ServiceNow Debuts Apriel Voice 27B, an Open Speech Model

ServiceNow has released a preview of Apriel Voice 27B, an open speech-language model that aims to handle spoken input and output within a single architecture rather than stitching together separate components. According to the project's paper listing on Hugging Face, the model is built around speech, reasoning, and tool use — the kind of capabilities that an AI assistant needs to act on voice commands, not just transcribe them.

At 27 billion parameters and a dense (non-MoE) design, Apriel Voice sits in a practical middle tier: large enough to carry reasoning ability, but small enough to be realistic for self-hosting. ServiceNow frames it as a unified system covering both ends of the audio pipeline.

What it brings together

• Text-to-speech generation for natural spoken output
• Speech recognition for transcribing audio to text
• Reasoning to interpret requests and plan responses
• Tool use, allowing the model to call external functions

The "Preview" label signals that this is an early look rather than a finished product, and the model carries a custom license rather than a standard permissive one — a detail worth checking before any commercial deployment.

Why it matters

Most voice assistants today chain a transcription model, a language model, and a speech synthesizer together, which adds latency and complexity. A single model that reasons and uses tools while handling speech natively could simplify that stack. It also extends ServiceNow's growing Apriel family into the audio domain, giving developers an open alternative for building voice-driven agents.