ServiceNow has released Apriel Voice 27B Preview in preview, an open model that understands spoken requests, reasons over the conversation and replies with text or tool calls. It does not speak with its own voice: it produces text, and a separate synthesizer is needed for speech.
The design joins two existing pieces. A speech encoder derived from Whisper large-v3-turbo passes audio into Gemma 4 26B-A4B through a learned adapter, so audio, text, images and tool results can share one conversation.
Image: ServiceNow-AI — https://huggingface.co/ServiceNow-AI/Apriel-Voice-27B-preview
Training happened in two stages: the speech interface is aligned first, then adapted together with the base model, mixing speech tasks with text replay to avoid losing earlier abilities. Recordings longer than 30 seconds are split automatically. According to ServiceNow's tests, continuous audio works well up to about half an hour, and thinking mode is recommended for longer input.
On English tasks the company reports 51.27% on EVA-A, 74.13% on EVA-X and 65.51% on τ-Voice, with a mean latency of 1.91 seconds. These are self-reported figures. Closed systems such as GPT Realtime 2.1 and Gemini 3.8 Live score higher on τ-Voice, but they are proprietary speech-to-speech products.
Image: ServiceNow-AI — https://huggingface.co/ServiceNow-AI/Apriel-Voice-27B-preview
For speech recognition in English, the word error rate is 2.23% on LibriSpeech test-clean and 5.15% on test-other. A vLLM image is provided for serving on four GPUs with an OpenAI-compatible API. The model is available on Hugging Face.
No comments yet. Be the first!