ServiceNow's Apriel Voice 27B Unifies Speech, Reasoning, and Tool Calls in One Open Model
The 27B open-weight model routes speech through the same backbone used for reasoning and tool-call generation
ServiceNow Research published a paper on November 2 describing Apriel Voice 27B Preview, a 27-billion-parameter open model that accepts spoken audio, performs multi-step reasoning, and executes function calls inside a single architecture. The model is built on Google DeepMind's Gemma 4 26B-A4B base and is available on Hugging Face under an MIT license, making it the first open-weight system in its capability class to directly compete against proprietary voice models from OpenAI and xAI.
One Architecture Replaces a Three-Step Pipeline — and Why That Matters
Most enterprise voice agent deployments today run three separate models in sequence: an automatic speech recognition model transcribes audio to text, a large language model reasons over the transcript and decides what tools to invoke, and a text-to-speech model converts the response back to audio. Each hand-off introduces two failure modes — errors that compound and latency that stacks.
When a speech recognition system produces an incorrect transcription, that error propagates into the LLM as a factual premise; when the LLM generates a function call, any ambiguity compounds again at the TTS stage. Across three models, even modest per-step error rates multiply into substantially higher end-to-end failure rates before the first user hears a word.
Apriel Voice 27B eliminates those hand-off points. Speech tokens and text tokens pass through the same model weights in the same forward pass, with no serialization step between understanding and deciding. ServiceNow's paper reports the model achieves 65.51% task completion on the τ-Voice benchmark — a full-duplex voice agent evaluation developed by Sierra AI that scores completions against actual database state changes rather than model self-assessments.
How the Model Is Trained Without Losing Its Reasoning Capability
The architecture begins with Gemma 4 26B-A4B, a mixture-of-experts model from Google DeepMind that carries 25.2 billion total parameters but activates only approximately 3.8 billion per token. Of its 128 specialized expert blocks per layer, each forward pass routes tokens through 8 active experts — giving the model large effective capacity at relatively low inference compute. The 256,000-token context window was a key selection criterion for long voice task sessions.
Training proceeds in two stages. Stage 1 freezes the Gemma backbone and trains only a Whisper large-v3-turbo speech encoder and a projector layer that maps audio representations into the LLM's embedding space. This establishes the speech interface without disrupting the reasoning capability the base model already holds.
Stage 2 is where Apriel Voice diverges from standard fine-tuning. The projector and the full backbone are jointly adapted using four simultaneous objectives: supervised fine-tuning on speech-grounded tasks, text replay to prevent catastrophic forgetting of the backbone's original text capabilities, source-dependent cross-entropy weighting that applies different supervision signals to speech and text inputs, and forward KL divergence alignment that continuously pulls the speech model's output distribution toward the original text model acting as an online teacher.
The online distillation setup is implemented for efficiency: the teacher model runs locally on the same hardware, its inference is overlapped with student gradient updates to minimize wall-clock training time, and the KL divergence is computed over full vocabularies using fused loss functions. The result is a model that adds speech understanding without reverting to the language quality of a smaller or less capable baseline.
Benchmark Performance and What the Numbers Actually Mean
Apriel Voice 27B Preview reports company-claimed scores across four evaluations. On Big Bench Audio, an audio reasoning benchmark covering 12 tasks including emotion recognition, spoken arithmetic, and factual question answering, the model scores 98.90 in thinking mode — above the 96.6% that OpenAI's GPT-Realtime-2 reports on the same benchmark, though neither result has been independently reproduced.
On Speech-BFCL, an audio adaptation of the Berkeley Function Calling Leaderboard developed by Salesforce AI Research and UC Berkeley containing 6,200 expert-verified function-calling tasks, Apriel Voice 27B scores 81.13. The τ-Voice score of 65.51% across 278 customer service voice tasks places it within approximately 1.8 percentage points of xAI's Grok Voice Think Fast 1.0, which scores approximately 67% on the same benchmark.
The mean end-to-end latency of 1.91 seconds on τ-Voice tasks is company-reported and deserves context. For enterprise use cases — scheduling, service requests, database queries — a two-second response latency is generally acceptable. For conversational dialogue that users expect to feel phone-like, it may not be. The preview designation signals that this is a research release, not a hardened production system, and latency optimization is a common focus between a research preview and production deployment.
Open Weights Against Closed Competitors
The competitive picture turns significantly on licensing. GPT-Realtime-2 is a closed API priced at $32 per million input tokens and $64 per million output tokens — enterprise voice agents processing thousands of conversations per day face meaningful per-conversation compute costs at those rates. Grok Voice Think Fast 1.0 runs at $0.05 per minute, a pricing model more aligned with telephony infrastructure but still fully proprietary.
Apriel Voice 27B's MIT license permits organizations to run inference on their own infrastructure at no per-token cost, keeping customer voice data on hardware they control — a material distinction for enterprises in regulated industries. The permissive MIT terms place no restriction on commercial redistribution or derivative products, giving enterprises and technology vendors broad latitude to build on or package the model as they choose.
The model extends a ServiceNow–NVIDIA collaboration that produced Apriel Nemotron 15B in mid-2025. The Apriel series began as a 5-billion-parameter enterprise reasoning model in April 2025; the 27B voice release is its first speech extension.
Limitations and the Road to Production
The paper's "Preview" label is not cosmetic. No independent evaluation of these benchmark scores has been published yet. Readers evaluating these results should treat them as company-reported until third-party researchers reproduce them.
The current ServiceNow production pipeline for voice — which pairs Ink-Whisper for speech recognition with Cartesia Sonic-2 for synthesis — is not replaced by Apriel Voice 27B Preview. The paper is a research contribution, and the path from a research preview to a platform-integrated production system typically involves additional safety evaluation, latency optimization, and enterprise-hardening that the paper does not address.
What the paper does establish is that the architectural premise is viable at this scale and with this training methodology. The remaining question is whether the end-to-end accuracy and latency profile, once independently verified, justify the operational complexity of running a 27B-parameter MoE model in-house rather than routing to a proprietary API. The open MIT license means the benchmark results, when independently confirmed, will be the deciding argument — not pricing negotiations.