Open-source voice agents have been stuck in the single digits — in ServiceNow's own evaluation table, UltraVox v0.6 scores just 8.94 on service accuracy and Audex-30B-A3B only 5.66. On Oct 29, ServiceNow put Apriel Voice 27B Preview (paper) on Hugging Face and pulled the open camp straight to 51.27. The pitch in one sentence: it takes audio in and returns text replies plus structured tool calls.

Two stages: it listens, but does not speak

Today's voice assistants mostly follow one of two routes: stitch ASR + LLM + TTS together, or go end-to-end speech-to-speech (S2S). Apriel Voice takes the middle path — two stages. A Whisper large-v3-turbo speech encoder connects, through a learned projector, to the post-trained Gemma 4 26B-A4B backbone. The model ingests 16kHz audio and outputs text and tool calls; speech synthesis is left to an external component. In the official evaluation all two- and three-stage systems use Cartesia Sonic 3.5, while proprietary S2S systems use their own. The model card says it plainly: the checkpoint itself does not do full-duplex speech.

The bet behind two stages is explicit: the speech interface only listens, while reasoning and tool calling reuse an existing open backbone. Speech capability can then ride backbone upgrades instead of retraining a whole model per modality; the price is that speaking forever depends on an external synthesizer, with latency and voice out of the model's hands.

Training runs in two stages. Stage one establishes the speech interface; stage two jointly adapts the projector and backbone with supervised fine-tuning, text replay, and source-dependent cross-entropy plus forward KL supervision. Conversational data comes from automatically constructed speech-conditioned dialogues, and distillation is online — local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses.

The score sheet: experience leads, task success lags

In thinking mode: 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, 65.51% task success on τ-Voice with 1.91 seconds mean latency, and a mean of 85.84 across eight audio-reasoning benchmarks. On transcription, LibriSpeech word error rate is 2.23% on test-clean and 5.15% on test-other, with GigaSpeech at 10.20, μ-Bench (en-US) at 10.76, FLEURS (en-US) at 3.55 and Earnings21 at 11.25 — lower is better.

Against proprietary S2S: GPT Realtime 2.1 sits at 59.03 service accuracy, 61.6 experience, 69.82 τ-Voice, 1.55 seconds latency; Gemini-3.8 Live Extended Thinking at 54.82, 73.5, 84.19, 1.50 seconds. In this official table, Apriel's 74.13 experience score is the highest of any system — 0.63 above Gemini and 12.53 above GPT Realtime. But its 65.51 task success trails Gemini by 18.68 points, service accuracy still sits 7.76 behind GPT Realtime, and latency is 0.41 seconds slower. Service accuracy measures whether the booking is right; experience measures whether the conversation flows. The gap between the two says it plainly: strong on conversational rhythm, execution still lagging. It converses well; the job is not yet finished.

Speech came in; text ability mostly stayed

Comparing before and after speech training: MMLU-Pro 85.27→84.42 (slightly down), GPQA Diamond 77.80→80.40 (up), AIME 2024/2025 unchanged at 92.00/81.67, MATH-500 92.76→92.88, and the τ² agent benchmark (mean across three domains, with GPT-5.2 as the user simulator) 68.93→70.95. Text replay clearly did its job — the speech interface did not eat the reasoning backbone.

Deployment and license

The checkpoint is 27B params in BF16 under an MIT license. ServiceNow ships a vLLM v0.32.0 Docker image; the reference setup runs on four H100 80GB GPUs (tensor-parallel 4) with a 262144 context window, and speculative decoding uses gemma-4-26B-A4B-it-assistant as the draft model. On four H100s the optimized image delivers roughly 2x the measured output throughput of the baseline configuration. Long audio is split into 30-second chunks, with good performance reported up to about 30 minutes; one-hour input requires raising the admission limit to 3600 seconds and the batch budget to 32768. The repo shows 338 downloads in the past month, and the author himself submitted the paper to the community board only today — it currently has 16 upvotes. This model is very new.

The cold water, from the vendor itself

The model card's Limitations section does not hold back: multilingual performance is uneven, and English results do not imply other languages; noisy and distant-microphone speech remain challenging; faithfulness and reliable completion of complex tool workflows are still limited; longer recordings can repeat themselves or lose coherence; and the benchmark scores describe the evaluated configurations only. Every number also comes from a single self-reported evaluation by ServiceNow, with no third-party replication yet — the "Preview" suffix means what it says.

The open camp now stands in the 50-point band on this evaluation table, validating the two-stage route of encoder plus open reasoning backbone plus projector. But an 84.19 task success rate and 1.5-second latency remain the proprietary moat. Once listening is solved, the next battlefield is getting the job done — a voice assistant that converses beautifully but fails to book the flight still scores zero.