Back home

Apriel Voice 27B

x 0.78
Visit source

Intro

Apriel Voice 27B (Apriel Voice 27B Preview) is an open-weight voice model from ServiceNow, released on October 29, 2026 with a technical report posted on November 2, 2026, under the MIT license. It understands spoken requests, reasons over conversational context and produces text responses plus structured tool calls, while keeping text and image input.

Architecturally, a Whisper large-v3-turbo speech encoder is connected through a learned projector to a Gemma 4 26B-A4B backbone, so one shared model handles listening, reasoning and tool use. Pair it with any external TTS to build voice assistants, customer-support flows and spoken tool-use apps.

Explain

Capabilities

Capability Details
Speech understanding Transcription, speech translation, spoken QA across ~50 languages
Reasoning & tools Instruction following, multi-step requests, structured tool calls
Multimodal context Audio, text, images and tool results in one conversation
Long audio Auto-chunks into 30 s segments; good results up to ~30 min recordings

Results (English, vendor-reported)

  • EVA-A 51.27%, EVA-X 74.13%, τ-Voice task success 65.51%, mean latency 1.91 s.
  • LibriSpeech WER: 2.23% test-clean, 5.15% test-other.
  • Clearly ahead of open three-stage "ASR + Gemma 4-26B" cascades on EVA, approaching some proprietary realtime models.

Why it matters

Most voice agents are cascades (speech-to-text → LLM → text-to-speech), so errors and latency stack up. Apriel Voice merges understanding and reasoning into one backbone and is MIT-licensed for private deployment. Note it does not generate speech itself and is not natively full-duplex; realtime turn-taking is handled by the app and serving stack.

Usage

Getting started

  1. Try it online — the Apriel Live Space on Hugging Face (servicenow-ai-apriel.hf.space) demos voice chat and tool calls.
  2. Download weights — huggingface-cli download ServiceNow-AI/Apriel-Voice-27B-preview, load with transformers (pipeline_tag: audio-text-to-text).
  3. Serve it — the model card includes serving examples (e.g. vLLM) with settings for 30-minute and optional one-hour audio; input is 16 kHz mono.
  4. Build a voice agent — your app executes tools and returns results to the conversation, then pipes text through any TTS engine.
  5. Long audio — enable thinking mode for longer recordings.

Needs substantial GPU memory (27B-class, ~4B active MoE backbone). It is a preview release — evaluate before production use.