ServiceNow's speech language model: It attaches a Whisper large-v3-turbo speech encoder and projection layer to Gemma 4 26B-A4B, enabling it to directly listen to speech and output text responses and structured tool calls without first transcribing and passing to another dialogue model; it also retains text and image inputs. Training occurs in two stages: first aligning the speech interface, then jointly tuning the projection layer and backbone using supervised fine-tuning, text replay, and cross-entropy / forward KL distillation differentiated by data source. It only outputs text and tool calls; speech synthesis must be added separately for spoken output.
Likes 33
Reasons
- The authors report that in thinking mode, task success rates for Big Bench Audio 98.90, Speech-BFCL 81.13, and τ-Voice are 65.51%, with an average latency of 1.91 seconds; the model card table shows an EVA experience score higher than the listed open-source two-stage and three-stage solutions.
- The weights are MIT licensed, and the model card provides usage instructions for vLLM deployment (OpenAI-compatible interface, automatic tool selection) and Transformers, allowing direct application in voice customer service, dictation-based tool calling, and similar scenarios.
- It ranked fifth in likes on the day of its Hugging Face release (indicating attention only); the model card states that text reasoning and tool calling scores are roughly on par with Gemma 4 26B-A4B, indicating that adding speech capabilities does not significantly degrade text performance.
Usage
- Online demo: Apriel Live from the model card (servicenow-ai-apriel.hf.space).
- Deployment: Use the vLLM v0.32.0 image provided in the model card (or the upstream vLLM PR #64683), with the command
vllm serve --model ServiceNow-AI/Apriel-Voice-27B-preview --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4; the example is configured for four GPUs. - Invocation: In the OpenAI client messages, include {"type": "audio_url"} along with text instructions; pass tools for tool calling and read message.tool_calls; for transcription tasks, use temperature 0.4 and disable thinking.
- Audio must be 16 kHz mono; long recordings are automatically chunked into 30-second segments, and the model card states that performance is better for recordings under approximately 30 minutes.
Risks
- This is a Preview version with no independent code repository; the deployment image is hosted under a personal Docker account, and official vLLM support is still in a PR.
- High deployment barrier: The example requires four H100 80GB-class GPUs; the end-to-end experience also depends on external TTS and turn management, as the model itself is not full-duplex speech.
- The model card notes uneven multilingual performance, difficulties with noisy and far-field speech, and potential errors in complex tool workflows; benchmark scores are as reported by the official team.