Apriel Voice 27B Preview

Apriel Voice 27B Preview is an open speech-language model developed by ServiceNow that unifies speech recognition, complex conversational reasoning, and structured tool calling into a single inference pipeline. Instead of relying on a fragile multi-model cascade where an automatic speech recognition (ASR) engine transcribes audio and passes text downstream to a large language model, Apriel Voice ingests speech directly. The architecture connects OpenAI's Whisper large-v3-turbo speech encoder to a post-trained Gemma 4 26B-A4B backbone via a learned projector.

The checkpoint operates as a 2-stage speech-to-text and speech-to-action engine: audio enters the model, and the model outputs text reasoning, conversational dialogue, or structured JSON tool calls. It does not synthesize native output audio. To build a full duplex voice assistant, developers pair this checkpoint with a dedicated text-to-speech (TTS) engine and an application-layer turn-taking controller. Released under open weights on Hugging Face (MIT license), Apriel Voice represents a major step forward for developers seeking a competitive local AI model with 27B parameters in 2026 for voice agent workflows.

Because Apriel Voice retains the multimodal backbone of Gemma 4, it accepts audio, text, and visual inputs within a shared context window. On standardized audio benchmarks, the model demonstrates state-of-the-art capability among open systems: in thinking mode, it achieves 98.90 on Big Bench Audio, 81.13 on Speech-BFCL for audio function calling, and a 65.51% task completion rate on the tau-Voice benchmark with an average response latency of 1.91 seconds.

Apriel Voice 27B Preview balances parameter capacity against inference latency through a Mixture of Experts (MoE) design. The model houses approximately 27.2 billion total parameters, but only routes tokens through 4 billion active parameters during inference passes.

This structure delivers distinct computational advantages:

• MoE Execution Profile: While the full 27B parameter footprint must reside in system or video memory, the computational cost per token matches that of a compact 4B parameter model. Apriel Voice 27B Preview MoE efficiency translates directly into lower time-to-first-token (TTFT) and significantly higher decoding speeds compared to dense 27B models.
• Acoustic Front-End: Audio ingestion runs through Whisper large-v3-turbo. Incoming audio is standardized to 16 kHz mono and chunked into 30-second windows before routing through a learned projection layer into the Gemma 4 backbone.
• Context Budget and Audio Ingestion: Audio chunks, text tokens, and image tokens share the backbone's unified context window. The processor supports continuous audio streams up to 30 minutes reliably, though longer inputs will consume a substantial portion of the KV cache.
• Thinking Mode: The model features an extended reasoning mode that performs chain-of-thought analysis prior to generating final text or tool payloads, which dramatically reduces hallucination rates during complex function calling.

Apriel Voice is built specifically for practical agentic workflows that require direct spoken comprehension rather than pure audio transcription.

Traditional voice agents fail when a user's spoken syntax is noisy, ambiguous, or filled with false starts. Because Apriel Voice was trained with joint cross-entropy and teacher-distillation supervision on spoken dialogues, it extracts intent directly from raw audio. Scoring 81.13 on Speech-BFCL, it reliably parses spoken commands directly into structured API parameters without intermediate transcription cleanup.

Because the Gemma 4 backbone retains its image ingestion capabilities, Apriel Voice can handle simultaneous speech and image inputs. A field technician can point a camera at machinery, describe a fault over audio, and receive immediate structured troubleshooting instructions or automated database lookup queries.

On standard ASR benchmarks, Apriel Voice achieves a Word Error Rate (WER) of 2.23% on LibriSpeech clean and 5.15% on LibriSpeech other. It natively supports translation and transcription across more than 50 languages, handling multilingual queries without requiring an upfront language-routing model.

Deploying Apriel Voice locally requires balancing two distinct hardware constraints: having enough total memory to store the full 27B parameter MoE structure, and providing enough compute bandwidth to process the Whisper audio encoder without adding latency to the conversation.

The total memory footprint is governed by the 27B MoE backbone plus the Whisper encoder weights and the active KV cache.

• BF16 / FP16 (Unquantized): Requires 56 GB to 62 GB of VRAM. Demands an A100/H100 (80GB) or a multi-GPU rig (such as two RTX 3090/4090 GPUs running tensor parallelism).
• 8-bit Quantization (Q8_0): Requires approximately 32 GB of VRAM. Fits onto a dual-GPU setup or an Apple Silicon machine with 48 GB+ unified memory.
• 4-bit Quantization (Q4_K_M): Requires 16 GB to 18 GB of VRAM. Fits entirely within a single 24GB consumer graphics card, leaving adequate headroom for context tokens and audio feature caches.

The best GPU for Apriel Voice 27B Preview at consumer tier is an Nvidia GeForce RTX 3090 or RTX 4090 (24GB VRAM). When running a 4-bit quant, a single 24GB card accommodates the entire model weight set, the Whisper audio processor, and a working context window without spilling over into system RAM.

For Apple Silicon users, an M3 Max or M4 Max with 36GB, 48GB, or 64GB of unified memory offers an ideal unified workstation environment, executing both the speech encoder and the MoE backbone locally with zero PCI bus bottlenecks.

Because only 4B parameters are active during token generation, Apriel Voice 27B Preview tokens per second run exceptionally fast once the initial audio prompt is ingested:

• RTX 4090 (Q4_K_M): 45 to 65 tokens per second during generation.
• Apple Silicon M4 Max (Q4_K_M): 35 to 50 tokens per second.
• Apple Silicon M3 / M4 (Unified Memory, CPU/GPU shared): 20 to 35 tokens per second.

The primary latency factor is not the generation speed, but the initial audio ingestion pass. The Whisper large-v3-turbo encoder processes a 10-second audio clip in roughly 150 to 300 milliseconds on modern tensor cores before text generation begins.

The quickest way to run Apriel Voice 27B Preview locally is via Hugging Face Transformers or vLLM once official pipeline support lands. For local deployment environments using quantization backends:

1. Quantization Selection: The best quantization for Apriel Voice 27B Preview is Q4_K_M or AWQ 4-bit. Given that the base model has 27B parameters, 4-bit weight compression incurs virtually zero degradation on reasoning or Speech-BFCL benchmark metrics while lowering the memory barrier to a single consumer GPU.
2. Audio Setup: Audio must be sampled or resampled to 16,000 Hz mono before feeding it to the pipeline processor.
3. Serving Stack: To implement real-time dialogue, connect the model's output stream to a fast local TTS engine such as Piper, Kokoro, or XTTS-v2. Application logic must monitor turn-taking, interrupting generation if a new voice activity detection (VAD) trigger indicates the user is speaking again.

Evaluating Apriel Voice against contemporary open-weight alternatives highlights the trade-offs between cascaded systems and unified speech-language models.

The standard baseline for local voice applications is running Whisper large-v3 independently, feeding the text output into a model like Gemma 2 27B or Llama 3 8B.

• Latency: Cascaded systems introduce an unavoidable double-inference penalty: ASR transcription must fully complete before the LLM can generate its first token. Apriel Voice achieves a mean latency of 1.91 seconds on tau-Voice, significantly outperforming cascaded setups that typically require 2.9 to 3.5 seconds.
• Acoustic Nuance: Cascades strip out inflection, pauses, emotional tone, and ambient acoustic cues during the text conversion step. Apriel Voice maintains acoustic embeddings directly into its reasoning layer.

Against other unified open models like UltraVox v0.6 (Qwen-based) or Audex-30B-A3B:

• Tool Calling Precision: While UltraVox and Audex deliver competitive raw transcription and basic conversational answers, both struggle significantly with complex multi-turn logic and API execution. On the EVA-A service accuracy benchmark, Apriel Voice posts 51.27%, compared to 8.94% for UltraVox v0.6 and 5.66% for Audex-30B-A3B.
• Conversational Experience: Apriel Voice reaches 74.13% on EVA-X (conversational experience), outperforming smaller open multimodal audio checkpoints by a wide margin and rivaling proprietary systems like Gemini Live.

If your use case strictly requires native voice generation from the model itself without an external TTS engine, you will need a full speech-to-speech architecture. But for developers building robust, local automated agents that take spoken commands, execute database tools, and respond with minimum latency on consumer hardware, Apriel Voice 27B Preview is currently the strongest open-weights model in its weight class.