Open Apriel Voice 27B Scores 98.90 on Big Bench Audio and 81.13 on Speech-BFCL

Researchers on the Apriel project have introduced Apriel Voice 27B Preview, an open speech language model that combines voice understanding, reasoning and tool calling in a single shared backbone. In reasoning mode, it scores 98.90 on Big Bench Audio (a set of reasoning questions delivered as audio) and 81.13 on Speech-BFCL (a speech version of the Berkeley Function-Calling Leaderboard benchmark). Average response latency is 1.91 seconds. The model is built on a fine-tuned checkpoint of Gemma 4 26B-A4B, a mixture-of-experts architecture with roughly 26B total parameters and about 4B active per token. Apriel Voice accepts speech directly, keeps its text and image inputs, and generates both responses and tool calls itself, so no separate dialogue model is needed.

What the benchmarks showed

Beyond the two headline results, the average task success rate on the Voice benchmark is 65.51%. On the English EVA-Bench, which measures both service accuracy and interaction quality, accuracy reaches 51.27% and the experience score pass@1 is 74.13%. The authors call this a strong result for an open model whose speech abilities extend to reasoning and action.

The authors also stress that recognition accuracy alone is not enough for a voice agent: a model can transcribe a request correctly and still fail to follow constraints or call the right tool. So they evaluated speech as an interface to general capabilities (reasoning, instruction following, dialogue and task completion) alongside recognition and translation. The report also compares modes with and without reasoning, language coverage, and how well the model retains its text and image skills.

Architecture: encoder, projector and shared backbone

Audio passes through a frozen Whisper large-v3-turbo encoder, which outputs 1280-dimensional acoustic features. A trainable projector stacks eight adjacent frames and runs them through linear layers with a SwiGLU gate into the decoder's 2816-dimensional space. The design is borrowed from Ultravox, and the projector has exactly 47,722,496 parameters, about 47.7M. No transcript is produced at inference: the decoder receives the acoustic representations directly.

Tool execution and speech synthesis sit outside the model. The application runs the tools, returns the results to the same context, and sends the response text to a swappable speech synthesizer. End-of-turn detection and interruption handling also run around the model. As a result, the voice or the synthesis system can be changed without retraining the reasoning model itself.

How the model was trained

The goal was to give an already strong decoder a speech interface without losing its existing capabilities. In the first stage, the projector is trained on multilingual transcription, translation and speech continuation while the backbone stays frozen. In the second stage, the projector and backbone are updated together, using speech tasks, internally generated dialogues, tool-use examples and replayed text data.

For each data source, the training combines cross-entropy with forward Kullback–Leibler divergence. The first teaches target responses and tool calls; the second transfers the behavior of a frozen copy of the original backbone and limits forgetting. This teacher is needed only during training and is absent at inference.

Dialogues are first created as text, then automatically converted to speech. The team selects fragments, uses a single reference voice within each dialogue, and checks conversion accuracy. The engineering work includes running the teacher locally, overlapping teacher and student computation, fused kernels for full-vocabulary losses, and consistent checkpoint saving.

What this means

For practitioners, the recipe itself is the interesting part: take a strong text model, add a speech projector, and preserve its behavior through distillation instead of building a cascade of speech recognition, an LLM and speech synthesis. This is a Preview release, and for live conversations the authors have deployed their own platform, Apriel Live.