ServiceNow releases Apriel Voice 27B Preview, an open model that listens, reasons and calls tools

Most voice assistants chain a speech recogniser into a separate text model. Apriel Voice instead connects a Whisper large-v3-turbo speech encoder to the post-trained Gemma 4 26B-A4B checkpoint through a learned projector, so audio, text, images and tool results all sit in the same conversation context[2]. ServiceNow's example is someone reporting a locked account who speaks a code, adds a screenshot and then changes their mind about what should happen next; the model has to connect those details and pick the right action, not just transcribe the sentence[1].

The model produces text and tool calls, not audio. Speech synthesis is a separate, swappable component, which in Apriel Live is Cartesia Sonic 3.6, so teams can change voices or providers without retraining the reasoning model[1]. Turn-taking is also handled by the application, and the checkpoint does not do native full-duplex speech[2].

In thinking mode ServiceNow reports 65.51% task success on τ-Voice with a mean response latency of 1.91 seconds, 51.27% accuracy and 74.13% experience on EVA, 98.90 on Big Bench Audio and 81.13 on Speech-BFCL[3]. On LibriSpeech, word error rate is 2.23% on test-clean and 5.15% on test-other[1]. Training happens in two stages: first aligning the speech interface, then adapting the projector and backbone together with supervised fine-tuning, text replay and online distillation from a teacher model[3].

Longer audio is split into 30-second chunks, and ServiceNow says it saw good results on continuous recordings up to about 30 minutes[2]. For serving, it ships a vLLM image whose reference command runs the model on four GPUs behind an OpenAI-compatible API[2].

Enterprise voice agents for IT help desks and customer support are exactly ServiceNow's market, and an MIT-licensed model that can hear a request and call a tool directly gives builders an alternative to closed realtime voice APIs. ServiceNow is candid that this is a preview: multilingual quality is uneven, noisy and distant-microphone audio still needs work, and faithfulness in complex workflows is a priority to improve[1].