Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use

Apriel Voice 27B Preview. An Open Model for Speech, Reasoning, and Tool Use. By Aman Tiwari and thirty others, from ServiceNow. Published on the twenty-ninth of October.

Recognising words is only the start of what a capable voice assistant has to do. It must understand the request, reason over context, keep a conversation going, and act through tools. Hearing a sentence correctly is of little use if the system cannot then work out what was meant, carry it across several turns, and do something about it. Those demands are usually met by different pieces, with speech handled in one place and dialogue in another.

Apriel Voice twenty-seven B Preview is an open speech-language model that brings those capabilities into one shared reasoning backbone, while keeping text and image input as well. It is built on the post-trained Gemma 4 26B-A4B checkpoint. The model processes speech directly and generates both responses and tool calls itself, with no separate dialogue model required. Speech goes in, and what comes out is either an answer or an action, produced by the same backbone that does the reasoning.

Adapting the model to speech takes two stages. The first establishes the speech interface. The second jointly adapts the projector and the backbone, using supervised fine-tuning, text replay, and supervision that combines source-dependent cross-entropy with forward Kullback Leibler divergence. Conversational training data comes from speech-conditioned dialogues that are constructed automatically.

Training also leans on online distillation, with a teacher model and a student model working at the same time. Three measures are what make that efficient: the teacher is served locally, teacher and student computation are overlapped rather than run one after the other, and the losses over the full vocabulary are fused.

Results are reported in thinking mode. Apriel Voice scores ninety-eight point nine on Big Bench Audio and eighty-one point one three on Speech-BFCL. On tau-Voice it reaches a task success rate of sixty-five point five one percent, with a mean latency of one point nine one seconds.

Beyond those, it reaches fifty-one point two seven percent English EVA accuracy, and seventy-four point one three percent on experience pass at one. The model has also been deployed on a custom serving platform, to support interactive voice use.

Less a pipeline of parts than a single model that listens, thinks and calls tools. Understanding, conversation and tool use are asked to share one backbone, which keeps its text and image inputs while gaining speech. The two training stages are what teach that backbone to hear, and the distillation machinery is what makes teaching it affordable. The deployment on a serving platform is the last step: putting the result in front of people who talk to it.

Apriel Voice 27B Preview. An Open Model for Speech, Reasoning, and Tool Use. By Aman Tiwari and thirty others, from ServiceNow. Published on the twenty-ninth of October. This is the long version: the whole paper rather than the abstract.

Picture an employee whose account is locked. They describe the problem aloud, read out an asset identifier, correct it, and ask for a ticket. Getting the words right is the easy part. The useful outcome is a resolved request. The authors put it plainly: recognition accuracy is necessary but insufficient. A model can transcribe a request perfectly and still fail to follow its constraints or call the right tool. The conventional answer is a cascade: a recognizer writes a transcript, a language model reasons over it, and a synthesizer speaks the reply. The reasoner never hears the speech, and every handoff costs time. ServiceNow's answer is Apriel Voice, one open model that hears speech directly and decides both what to say and which tool to call.

Two things matter before the mechanism. The first is what is kept. The model starts from a post-trained Gemma 4 checkpoint, a mixture of experts with about twenty-six billion parameters, of which about four billion are active for each token. It already reads text and images, reasons, and calls tools. The job is to add an ear without breaking any of that. The second is what is left outside. Speech synthesis is not part of the model: response text goes to a separate synthesizer that can be swapped without retraining the reasoner. Turn detection and interruption control also sit in the application. This is a speech-conditioned reasoner, not a speech-to-speech model.

The ear is a frozen Whisper large-v3-turbo encoder. Its acoustic frames pass to a small projector built the way Ultravox builds one: eight adjacent frames are stacked into one, shortening the sequence eightfold, then projected into the decoder's embedding space. The projector has about forty-seven point seven million parameters. Its outputs replace reserved audio positions in the input, so spans of speech can sit between pieces of text. No transcript is ever produced.

Stage one trains only the projector, on transcription, translation and continuation, over about thirty-two thousand hours of audio. Stage two also trains the backbone and its expert router, on mathematics, conversation, tool use and replayed text.

The conversations are synthetic. The team writes text dialogues with requests, answers, tool calls and results, and a selector chooses which parts of each user turn to speak. It favours keeping code, tables and case-sensitive identifiers as text, so the asset number stays written while the complaint is spoken. Every converted dialogue keeps its text twin, and that twin is what makes the supervision work.

Each data source chooses its own loss. Cross-entropy teaches target answers. Forward Kullback–Leibler divergence pulls the student toward a frozen copy of the original Gemma. The student hears the audio version while the teacher reads the text version. Their inputs differ in length, but both predict the same response, so the distributions are compared only over response positions. The paper shows that when both losses apply, their sum behaves like cross-entropy against one soft target: a blend of the correct token and the teacher's whole distribution, set by the ratio of the two weights.

Running that teacher online is costly. Each of eight nodes holds two teacher GPUs and six student GPUs. A one-batch lookahead asks the teacher about the next batch while the student trains on the current one. Stage two finished in about three days.

The headlines are in the abstract: ninety-eight point nine on Big Bench Audio, and sixty-five point five one percent task success on Tau-Voice. The comparison that tests the thesis most directly is against a cascade on the same backbone, the Parakeet recognizer feeding Gemma 4. On Tau-Voice that cascade succeeds on sixty-seven point seven eight percent of tasks, responding after three point zero one seconds. Apriel Voice responds after one point nine one seconds. Completion is roughly equal; the answer arrives over a second sooner.

EVA, which scores service accuracy and conversational experience, shows a larger gap. An experience pass requires concise responses, progress, and good turn-taking, all at once. The cascade passes three point nine three percent of the time; Apriel Voice, seventy-four point one three percent. But faithfulness to the service interaction falls from sixty-six point three three percent for the cascade to forty-three point four nine.

Against its own starting point, text ability mostly holds: MMLU-Pro slips by under a point and GPQA Diamond rises. Six of nine image scores fall.

There is a limitations section, and it is candid. The model prioritizes English; multilingual ability is assessed mainly through translation, and the authors call it uneven. Aggregate scores, they write, do not establish paired significance or non-inferiority to the starting backbone. Synthetic data and distillation can propagate wrong targets or wrong teacher predictions, and the fidelity checks confirm that audio matches text, not that an answer is right. There are no ablations, so nothing can be credited to any single ingredient. Transfer to other backbones is untested. And the voice scores include an external synthesizer, so they describe the complete system.

The idea to carry away is that giving a capable model an ear is mostly a problem of not losing what it already knew. The audio path is small. The care goes into supervision: every spoken example keeps a written twin, so the original model can read the text and show the adapted one, hearing the speech, what it would have said. Where that works, one backbone matches a cascade at finishing tasks while answering sooner. Where it frays, in faithfulness and some image scores, the paper's own tables say so.