AI BriefingServiceNow Releases Apriel Voice 27B Preview: Open Model for Speech, Reasoning, and Tool Use
·2026.10.30 01:53
Key point
The open-source model uses a Gemma 4 backbone and Whisper encoder to handle speech, reasoning, and tool calls, outperforming other open-source systems on voice benchmarks.
Details
- ServiceNow released Apriel Voice 27B Preview, an open model combining speech understanding, reasoning, and tool use via a shared Gemma 4 backbone.
- The model achieves 51.27% EVA-A, 74.13% EVA-X, and 65.51% τ-Voice task success with 1.91 seconds mean response latency.
- It supports transcription, speech translation, and spoken instruction following, with a word error rate of 2.23% on LibriSpeech test-clean.
- The architecture connects a Whisper large-v3-turbo speech encoder to the Gemma 4 26B-A4B backbone through a learned projector.
- Real-time voice interaction is managed by the application and serving stack, as the checkpoint generates text and tool calls rather than native full-duplex speech.
- Long audio recordings up to 30 minutes are supported, with performance potentially declining for longer durations.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
BottleCap releases ThinkingCap for Gemma-4-26B-A4B, cutting thinking tokens by 39% on average
Gemini 4 Carbon early access begins with improved 3D performance and slower speed