Apriel Voice 27B Preview: ServiceNow’s Open MIT Speech Model Built on Gemma 4

ServiceNow-AI has released Apriel Voice 27B Preview, an open-weight model that takes spoken audio and answers with text and tool calls. It carries the MIT license, sits on top of Google's Gemma 4 26B-A4B, and adds a Whisper large-v3-turbo speech encoder through a learned projector. The weights are on Hugging Face, and the technical report is on arXiv and the Hugging Face papers page. It is a preview, and it does not speak back: you need a separate text-to-speech system to build a full voice assistant.

This post covers what the model is, what the authors report, how it compares with the closed voice systems it is measured against, how to serve it, and where the limits are.

TL;DR: questions builders ask first

What is Apriel Voice, and how is it built?

The design is simple to describe. A frozen Whisper large-v3-turbo encoder turns audio into 1280-dimension features. Eight adjacent frames are stacked, and a learned projector of about 47.7 million parameters maps them into the language model's embedding space. The projected audio replaces reserved positions in the input sequence. From there, the Gemma 4 backbone treats speech like any other input.

This matters for builders because it avoids a pipeline of three models (speech recognition, then a language model, then a tool router). One model hears the request, reasons, and emits a tool call. The paper says it does this "without a separate dialogue model". Text and image input are still supported.

Training happens in two stages, per the paper:

1. Interface learning. Only the projector is trained, on multilingual transcription, translation and continuation: 18.17 million examples across 1,164 dataset configurations.
2. Broad task learning. The projector, the backbone and the mixture-of-experts router are updated on math, science, conversation, speech tasks, tool use and text replay. Each data source uses cross-entropy, forward KL divergence against a frozen teacher, or a weighted mix.

Stage 2 used 16 teacher GPUs and 48 student H100 GPUs for about three days. Synthetic speech is made by converting text dialogues into speech-conditioned examples, with a consistent reference voice per conversation.

The base model is the same family covered in our Gemma 4 26B-A4B on Mac post, which is useful context if you plan to test locally.

What do the benchmarks show?

A microphone sending a sound wave to a speaker, showing how an open speech model like Apriel Voice handles audio

All numbers below come from the model card and paper, not from an independent test. The authors note that comparisons mix internal and public evaluation settings, so this is not a controlled ranking.

Voice system scores (English, thinking mode)

Read this table plainly. Apriel is behind on accuracy and task success and slower on latency. It leads GPT Realtime 2.1 on the experience score and roughly ties Gemini on it. For an open model you can host yourself, being in range of closed systems is the story, not winning.

The paper breaks tau-Voice into domains: airline 69.33%, retail 81.58% and telecom 45.61%. The telecom number is a warning. Multi-step tool workflows are where the model is weakest, and the card says so directly.

Text and reasoning retention

A common worry with adding speech is that the base model gets worse at text. The card compares against Gemma 4 26B-A4B in thinking mode:

Text skill is essentially preserved, and tool use is slightly better. Image retention is also reported: MMMU-Pro 69.88 against 70.23, and MathVision MINI 67.10 against 64.14.

Speech recognition and audio reasoning

On word error rate, lower is better: LibriSpeech test-clean 2.23, test-other 5.15, GigaSpeech 10.20, FLEURS en-US 3.55 and Earnings21 11.25. The paper reports an average recognition WER of 12.61 across its set. Speech translation (German, French and Spanish into English) averages 42.87 BLEU, close to GPT Realtime 2.1 at 43.72.

For spoken reasoning, the card lists Big Bench Audio 98.90, BFCL 81.13, IFEval 86.47, MT-Bench 86.19 and Spoken SQuAD 82.24, with an eight-benchmark audio reasoning mean of 85.84. The authors call that the strongest among open two-stage models.

How do I run it?

The model card gives a vLLM recipe. The standard setup uses a Docker image built on vLLM v0.32.0 (or upstream vLLM pull request 64683) and these key flags:

vllm serve ServiceNow-AI/Apriel-Voice-27B-preview \
  --served-model-name Apriel-Voice-27B \
  --tensor-parallel-size 4 \
  --max-model-len 262144 \
  --max-num-batched-tokens 16384 \
  --enable-auto-tool-choice \
  --tool-call-parser gemma4 \
  --reasoning-parser gemma4

That command is a summary of the flags on the card, not a tested copy of it; use the exact image names and flags from the model page. Notes from the card:

• The server exposes an OpenAI-compatible API on port 8000. Send audio through an audio_url content part with a text prompt.
• Toggle thinking through chat_template_kwargs. Use thinking for reasoning and workflow tasks, and turn it off for transcription.
• Tool calls arrive in message.tool_calls.
• For 30-minute audio set VLLM_MAX_AUDIO_DECODE_DURATION_S=1800. For one hour, use 3600 and a 32,768-token batch budget, memory permitting.
• An optimized image reports about twice the output-token throughput of the baseline on four H100 80 GB GPUs, using speculative decoding with a Gemma 4 26B-A4B draft model.
• With Transformers, load with AutoModel and AutoProcessor using trust_remote_code=True, and feed 16 kHz mono audio.

Suggested sampling: temperature 1.0, top-p 0.95 and top-k 64 for general instruction following. For recognition and translation use temperature 0.4 with the same top-p and top-k.

Because the model needs four GPUs in the reference setup, this is a server model, not a laptop one. If you want local speech tools on a small machine, see our guides to Kokoro local TTS and VoiceStudio.

What are the limits?

The model card is blunt, and builders should read it before shipping.

• No speech output. It generates text and tool calls only. Speech synthesis and turn handling are external, and it is not native full-duplex speech.
• English first. It is tagged for 49 languages, but the card says multilingual quality is uneven. English results do not imply the same quality elsewhere.
• Noisy audio. Distant microphones and noisy rooms remain hard. Audio should be 16 kHz mono.
• Long recordings. Repetition and loss of coherence can appear on very long audio. Thinking mode is recommended for longer clips.
• Faithfulness and tool workflows. Complex multi-step tool use is a weak point, as the telecom score suggests.
• Preview status. The name says it. Expect changes.
• Unlisted data. The model card does not list specific training datasets, so check the technical report if provenance matters to you.

For any action with side effects, such as refunds, account changes or bookings, keep a human or a policy layer between the model's tool call and the real system. Voice adds a new error source: a misheard word becomes a wrong argument in a tool call. Our coverage of OpenAI's GPT Live 1 voice agents API shows how closed vendors handle that with managed pipelines.

How does it fit the open voice landscape?

Open speech models are still thin compared with text models. Most open options handle either recognition or synthesis, not reasoning over speech with tools. Apriel Voice is notable because it merges recognition, reasoning and tool calling in one checkpoint under a permissive license.

Closed rivals are ahead on task success. xAI shipped Grok Voice Transcribe 2 and Meta shipped Muse Voice Transcribe for realtime speech to text, both as hosted services. The trade is the usual one: hosted systems are faster and more accurate today, while an MIT model lets you keep audio on your own servers, tune it, and audit it.

That privacy point is the strongest reason to try it. Voice recordings are sensitive. If your use case is a call center, a clinic or an internal helpdesk, running the model yourself avoids sending audio to a third party. ServiceNow's own business is enterprise workflows, so tool calling for tickets and requests is the obvious fit. See our post on ServiceNow's Arc agent for the company's agent direction. We do not claim a direct product link between Arc and this model; neither source says so.

Should you try it this week?

Try it if you:

• already serve vLLM models on a multi-GPU box,
• need speech-to-tool-call behaviour with audio kept in-house,
• can accept English-first quality and add your own text-to-speech.

Wait if you:

• need production-grade accuracy on multi-step tool tasks,
• need many languages at equal quality,
• want a full duplex, speak-and-listen model out of the box.

A sensible test plan: record 50 real requests from your own users, run them through Apriel and through your current hosted option, and compare the tool arguments, not just the transcripts. A model that transcribes well but fills the wrong field is worse than one with a slightly higher word error rate.

Related reading

• Gemma 4 26B-A4B on Mac with MLX
• OpenAI GPT Live 1 and voice agents
• Grok Voice Transcribe 2
• Meta Muse Voice Transcribe
• Kokoro local TTS guide
• VoiceStudio, an open source ElevenLabs alternative
• ServiceNow Arc agent

Specs and scores are as listed on the Hugging Face model card and paper on November 2, 2026, and may change as the preview is updated.