Description
Apriel Voice 27B Preview is an advanced open model developed by ServiceNow for speech, reasoning, and tool use. This audio-text-to-text model, featuring 27 billion parameters, is built upon the Transformers architecture and leverages a shared Gemma 4 backbone. It is designed to understand spoken requests, engage in reasoning over conversational context, and generate both text responses and structured tool calls.
Read More
The model significantly expands capabilities by integrating a speech interface, while also maintaining its original text and image input functionalities. This makes Apriel Voice 27B Preview highly versatile for developing interactive voice assistants, streamlining customer support workflows, and enabling spoken tool use. When paired with an external speech synthesizer, it can power real-time voice applications.
Key capabilities include comprehensive speech understanding, encompassing transcription, translation, and spoken question answering. It excels in reasoning and tool use, allowing it to follow complex instructions, handle multi-step requests, and generate structured tool calls. The model supports multimodal context, processing audio, text, images, and tool results within a shared conversational framework. For voice applications, it generates text responses that are then paired with external speech synthesis, with application-managed turn-taking.
Technically, Apriel Voice 27B Preview connects a Whisper large-v3-turbo speech encoder to the Gemma 4 26B-A4B backbone via a learned projector. Its training process involves aligning the speech interface and then adapting it with the backbone using speech tasks and text replay. The model supports long audio inputs, automatically splitting 16 kHz mono recordings into 30-second chunks, and has shown good performance on continuous recordings up to 30 minutes, with support for longer durations under specific conditions.
Performance metrics highlight its effectiveness, achieving 51.27% EVA-A, 74.13% EVA-X, and 65.51% τ-Voice task success, with a mean τ-Voice response latency of 1.91 seconds. On LibriSpeech, it demonstrates a word error rate of 2.23% on test-clean and 5.15% on test-other. The model is available for standard and optimized serving via Docker images, providing an OpenAI-compatible API for integration. It is released under an MIT License.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Apriel Voice 27B Preview Highlights
Audio-Text-to-Text Model: Processes spoken requests and generates text responses.
27 Billion Parameters: A large-scale model for complex tasks.
Transformers Architecture: Built with a robust and widely used deep learning framework.
Speech Understanding: Supports transcription, translation, and spoken question answering.
Reasoning and Tool Use: Follows instructions, handles multi-step requests, and generates structured tool calls.
Multimodal Context: Integrates audio, text, images, and tool results in conversations.
Voice Applications: Generates text responses for external speech synthesis in interactive voice assistants.
Gemma 4 Backbone: Utilizes a shared Gemma 4 26B-A4B backbone for core processing.
Whisper large-v3-turbo Encoder: Connects speech interface through a learned projector.
Long Audio Support: Automatically splits recordings into 30-second chunks, supporting up to 30 minutes continuously.
OpenAI-Compatible API: Provides an API for easy integration and inference.
Docker Images for Serving: Offers standard and optimized vLLM Docker images for deployment.
MIT License: Available for use under an open-source license.
High Performance Metrics: Achieves strong results in EVA-A, EVA-X, and τ-Voice task success.
Low Word Error Rate: Demonstrates competitive WER on English speech recognition benchmarks.
Getting Started with Apriel Voice 27B Preview
Access Model: Download or explore the Apriel Voice 27B Preview model files on Hugging Face.
Set Up Environment: Use the provided vLLM Docker image (e.g., `docker.io/amant555/vllm_apriel_voice:plain-26b-v0.32.0`) for serving.
Serve the Model: Run the Docker image on appropriate GPU hardware, exposing an OpenAI-compatible API on port 8000.
Configure Audio Duration: Adjust `VLLM_MAX_AUDIO_DECODE_DURATION_S` and `--max-num-batched-tokens` for desired audio input lengths.
Integrate via API: Send audio or text inputs to the served OpenAI-compatible API using the model name `Apriel-Voice-27B`.
Process Responses: Handle text responses and inspect `response.choices[0].message.tool_calls` for function calling outputs.
Pair with Synthesizer: Integrate an external speech synthesizer for real-time voice interaction in applications.
Optimize Performance: Utilize the optimized Docker image (`docker.io/amant555/vllm_apriel_voice:optimised-26b-v0.32.0`) for higher throughput.
Apriel Voice 27B Preview's Use Cases
- Interactive Voice Assistants
- Customer Support Automation
- Spoken Tool Use
- Speech Transcription
- Speech Translation
- Spoken Question Answering
- Multimodal Conversational AI
- Instruction Following
- Real-time Voice Interaction









