Apriel Voice 27B Preview: open speech model, explained

Apriel Voice 27B Preview is an open speech model from ServiceNow AI that listens to audio, reasons about it and calls tools, all in one network. The team describes it in a paper posted to arXiv on October 29, 2026. Most voice assistants chain a speech recognizer to a separate language model. Apriel Voice skips the transcript and feeds the audio straight into the language model.

It matters because it is a working open alternative to that chain, with weights under an MIT license. It also trails the best closed system in the paper's own tables, which makes it a useful test of how far open voice agents have come.

What Apriel Voice 27B does

The model takes speech as input and produces text: either an answer or a tool call, such as looking up a booking. It still accepts text and images, because it is built on Google's Gemma 4 26B-A4B, a mixture-of-experts model that runs only about 4B of its roughly 26B parameters for each token. We covered the license of that base family in Is Gemma 4 free?.

The model card says the checkpoint does not speak. It generates text and tool calls, so you need a separate speech synthesizer to talk back. It is also labeled a preview.

How the team built it

Three pieces are stitched together. A frozen Whisper large-v3-turbo encoder turns audio into features. A small learned projector of about 47.7M parameters maps those features into the language model's input space. The Gemma 4 backbone does the rest. No transcript is created along the way.

Training ran in two stages:

• Stage 1 trained only the projector, on 18.17M examples of transcription, translation and continuation drawn from public sets such as Common Voice, GigaSpeech, People's Speech, LibriSpeech and CoVoST 2.
• Stage 2 trained the projector, the backbone and the expert router together on math, science, conversation, noisy speech, tool use and synthetic dialogues. It took about three days on 8 nodes of H100 GPUs.

To keep the model from forgetting how to reason as text, the authors mixed in text-only data and trained against a frozen copy of the original model, using a forward KL loss, which pulls the new model's word probabilities toward the old one's. The synthetic dialogues were text conversations converted to speech with voice-cloning text-to-speech.

Results in the paper

All numbers below come from the authors' own evaluations, in thinking mode where the paper says so.

Against GPT-realtime-2.1, Apriel Voice averages 85.84 to 85.38 on the speech reasoning set, and it scores higher on Speech-BFCL (81.13 against 80.24). GPT-realtime-2.1 is ahead on τ-Voice (69.82%) and latency (1.55 s). The paper's EVA benchmark, which scores spoken customer-service conversations, gives Apriel Voice 51.27 on accuracy and 74.13 on experience.

Two comparisons stand out. Pipelines that pair a speech recognizer with Gemma 4 26B are slower: 2.92 to 3.01 seconds of latency against 1.91. And the best such pipeline, Parakeet plus Gemma 4 26B, still edges Apriel Voice on the speech reasoning average, 86.63 to 85.84. The advantage of one model is speed and turn-taking, not a higher score on every test.

On plain transcription, the average word error rate across eight sets is 12.61%, better than Flux (13.59%) and GPT-realtime-2.1 (13.56%) but behind Parakeet (10.06%). On LibriSpeech it scores 2.23% on the clean set and 5.15% on the harder one.

Speech-to-text translation averages 42.87 BLEU into English from German, French and Spanish, against 46.76 for Gemini 3.8 and 43.72 for GPT-realtime-2.1.

Did it lose its text and image skills?

Mostly not. In thinking mode it matches its base model on AIME 2024 (92.00), AIME 2025 (81.67) and AMC 2023 (100.00), and scores 80.40 on GPQA Diamond against 77.80 for Gemma 4 26B. On nine image benchmarks it wins three and loses six, usually by a point or two. In non-thinking mode the τ² agent benchmark drops from 53.23 to 49.14.

Limits the authors state

The paper says English is the priority and multilingual quality is uneven. Scores come from different prompts and judges than competing systems use, so they are not a controlled ranking. Synthetic training data and distillation can carry over errors from the teacher, and only one backbone family was tried. The model card adds that noisy and distant-microphone speech is hard, that complex tool workflows are not reliable, and that very long recordings can cause repetition.

Running it

The weights are on Hugging Face in BF16. The card shows Transformers and vLLM serving, with examples on four GPUs, and it was tested on four H100 80 GB cards. Audio is 16 kHz mono, split into 30-second chunks. The authors also run a demo, Apriel Live. We found no public code repository linked from the paper or its Hugging Face page, and the card notes upstream vLLM support through a pull request that was still open.

Next to watch: independent tests on τ-Voice and EVA, since every number above is the authors' own.