by ServiceNow
Open-weights 27B speech-to-text model from ServiceNow that adds a speech interface to a Gemma 4 backbone, with tool calling.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Apriel Voice 27B Preview is an open-weights model from ServiceNow that understands spoken requests and generates text responses and tool calls. It connects a Whisper large-v3-turbo speech encoder to a Gemma 4 26B-A4B backbone through a learned projector, and keeps the backbone's text and image input capabilities. The model card lists support for transcription, speech translation, spoken question answering, instruction following, multi-step reasoning and function calling.
The checkpoint outputs text and tool calls, not native speech. Voice assistants need an external speech synthesizer and application-managed turn-taking, and the application executes tools and returns results. Longer 16 kHz mono recordings are split into 30-second chunks. The authors report good performance up to about 30 minutes of audio, and recommend thinking mode for longer inputs.
The card reports 51.27% EVA-A, 74.13% EVA-X and 65.51% τ-Voice pass@1 with 1.91 s mean latency, plus LibriSpeech WER of 2.23% (test-clean) and 5.15% (test-other). These figures are the vendor's own, and it is labelled a preview. The weights are on Hugging Face under the MIT license and load via transformers with custom code. The card tags dozens of languages, including most EU languages.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-weights 27B voice-first multimodal MoE model that takes speech, text and images and acts on them, including text-to-speech output. It scores well for its size on benchmarks.
Closest neighbours by meaning, not just by category.
Model · Edge0
Open-weights 0.17B-parameter multilingual text-to-speech model with zero-shot voice cloning, runs on CPU via ONNX.
Model · Ant Group
Open-weights text-to-speech models from Ant Group's Inclusion AI with voice design, cloning and tag-based control.
Model
8.5 MB, 8M-parameter text-to-speech model distilled from Kokoro-82M, with four preset voices and CPU inference.
Model
Compact 0.6B open text-to-speech model with zero-shot voice cloning and support for 23 languages.
Model
Open-source append-only streaming speech recognition model with configurable low-latency decoding, English and Mandarin.
Model · Google
Google's open-weights 12B encoder-free multimodal model in the Gemma 4 family, taking text, image and audio input.