ServiceNow's Apriel Voice Bolts a Whisper Ear onto Gemma 4 and Releases It Under MIT

Most voice agents today are a relay race: one model transcribes the caller, another decides what to do, a third speaks. ServiceNow’s new open model collapses the first two legs into one, listening to audio directly and answering with text or tool calls. The company’s AI research team released Apriel Voice 27B Preview on October 29, 2026, with weights on Hugging Face under the MIT licence.

It targets enterprise voice agents, where a spoken request has to end in an action such as changing a booking. It is a preview, and the team is explicit about its gaps.

Gemma 4 with an audio front end

Apriel Voice is a speech-language model built in three parts. OpenAI’s Whisper large-v3-turbo encoder turns raw audio into acoustic representations, a learned projector maps those into the language model’s embedding space, and a post-trained Gemma 4 26B-A4B checkpoint does the reasoning. Text, images and tool results can share the same conversation context. The model outputs text and tool calls only. It does not produce speech, so a deployment still needs a separate text-to-speech system, and it is not full-duplex, meaning it cannot listen and talk at the same time.

Training ran in two stages. First, only the projector was trained, with the language model frozen, on multilingual transcription, speech translation and speech continuation, so that audio would land in a space the backbone already understood. Second, the projector and backbone were fine-tuned together on audio understanding tasks and voice-agent conversations. To stop the model losing its text skills, the second stage mixed in replayed text data and used a frozen copy of the starting backbone as a teacher, which is dropped at inference.

Scores against GPT Realtime and Gemini Live

ServiceNow evaluated the model on its own EVA framework, which scores voice agents on accuracy (EVA-A) and experience (EVA-X), and on τ-Voice, which tests spoken workflows across airline, retail and telecom. Apriel ran in thinking mode for both. In the comparison, the two- and three-stage systems used Cartesia’s Sonic 3.5 for speech, while the closed realtime models used their native voices.

In English, Apriel Voice scored 51.27 percent on EVA-A, 74.13 percent on EVA-X and 65.51 percent on τ-Voice, with a mean response latency of 1.91 seconds. GPT Realtime 2.1 posted 59.03, 61.6 and 69.82 percent at 1.55 seconds. Gemini 3.8 Live with extended thinking posted 54.82, 73.5 and 84.19 percent at 1.50 seconds. A conventional cascade of NVIDIA’s Parakeet speech recogniser feeding Gemma 4 26B scored 39.97, 3.93 and 67.78 percent and took 3.01 seconds to respond.

The picture is uneven across domains. On τ-Voice’s retail tasks, Apriel Voice scored 81.58 percent, the best result among the systems ServiceNow tested. Airline tasks came in at 69.33 percent, close to GPT Realtime 2.1’s 70 percent. Telecom troubleshooting was the weak spot at 45.61 percent, and the team names it as an area for improvement.

As a plain recogniser, the model posts word error rates of 2.23 on LibriSpeech test-clean and 5.15 on test-other. Across eight audio-reasoning benchmarks it averages 85.84.

The text model survived the surgery

Bolting audio onto a language model can erode its text ability, so ServiceNow published text scores against the base Gemma 4 26B-A4B. Apriel Voice scored 80.40 on GPQA Diamond against the base model’s 77.80, 70.95 on τ² against 68.93, and 84.42 on MMLU-Pro against 85.27.

Four H100s and a draft model

The recommended path is a vLLM 0.32.0 Docker image across four GPUs. ServiceNow’s optimised configuration adds speculative decoding with the Gemma 4 assistant model and is about twice as fast as the baseline: 1.96 times at a concurrency of one and 1.83 times at fifty, measured on four H100 80GB cards.

Long recordings are handled by splitting audio into 30-second chunks before encoding. The team says the model works well on inputs of up to about 30 minutes, which opens use cases such as summarising a recorded call, not just answering live ones.

ServiceNow’s own demo, Apriel Live, pairs the model with Cartesia’s Sonic 3.6 for the spoken reply.

Uneven languages and noisy rooms

The model card lists a long set of languages, but the team warns that multilingual performance is uneven and that English results do not imply comparable quality elsewhere. Background noise, distant microphones and meeting audio also leave substantial room for improvement, by ServiceNow’s own account. Combined with the missing speech output and half-duplex design, that makes Apriel Voice a building block for teams willing to assemble their own pipeline rather than a drop-in replacement for closed realtime APIs. What it offers in return is a permissively licensed model that can hear a caller and call a tool on its own hardware.