Open Apriel Voice 27B model scores 98.90 on Big Bench Audio speech test

Researchers on the Apriel project have released Apriel Voice 27B Preview, an open speech language model built on a fine-tuned Gemma 4 26B-A4B checkpoint. It takes voice input directly, reasons and calls tools without a separate dialogue model. In reasoning mode, the model scored 98.90 on Big Bench Audio, a test of reasoning over questions asked in audio format, and 81.13 on Speech-BFCL, which tests function calling from a spoken request. It also reached an average task success rate of 65.51% on a voice benchmark, with an average latency of 1.91 seconds.

One backbone instead of a cascade

A classic voice assistant is assembled from three parts: speech recognition (ASR), a language model and speech synthesis (TTS). Another approach leaves the voice layer with conversation only and hands reasoning and tools to a separate backend model. In Apriel Voice, a single model handles everything: it understands speech, replies in the dialogue and decides on its own when to call a tool and which one. Text, images and the results of tool calls stay in context.

The base model, Gemma 4 26B-A4B, is a mixture-of-experts model with about 26 billion parameters in total and about 4 billion active per token. The application executes the tools, returns the results to the conversation and passes the text of the reply to a swappable voice synthesizer, so the voice can be changed without retraining the model. Turn-end detection and interruption handling run around the model.

Speech reaches the decoder through a frozen Whisper large-v3-turbo encoder, which outputs acoustic features of dimension 1280, and a trainable projector. The projector concatenates every eight adjacent frames, passes them through linear layers and a SwiGLU block, and maps them into the decoder's embedding space of dimension 2816. It has exactly 47 722 496 parameters. No transcript is created at inference: the model works directly with acoustic representations.

Two training stages and online distillation

In the first stage, the backbone is frozen and only the projector is trained, on multilingual transcription, translation and speech continuation tasks. In the second stage, the projector and the backbone are fine-tuned together. This stage uses supervised fine-tuning on speech tasks, internally generated dialogues and tool interactions, plus replayed text data so the model does not lose its earlier skills.

The speech dialogues are built automatically. Text task dialogues are turned into spoken examples: the pipeline selects which passages to voice, uses the same reference voice throughout each conversation, and checks conversion accuracy separately.

Different data sources use different loss functions: cross-entropy, forward KL divergence (FKL) or a combination of the two. Cross-entropy teaches the target answers, while FKL against a frozen copy of the original backbone carries its behavior over to speech contexts. The teacher is needed only during training and is not present at inference.

To keep this setup computationally feasible, the authors run the teacher locally alongside the student, overlap teacher and student computation in time, and use fused loss kernels over the full vocabulary and coordinated checkpoint saving.

Results and context

Besides Big Bench Audio and Speech-BFCL, the authors evaluated the model on EVA-Bench, which measures both service accuracy and interaction quality. In the English version, accuracy was 51.27% and the experience score pass@1 was 74.13%. The report also covers multilingual coverage, the effect of reasoning mode, translation, enterprise scenarios and retention of text and image capabilities.

The method builds on known approaches: an encoder-plus-projector design, as in Ultravox and SLAM-ASR, and cross-modal behavior transfer, as in BLSP-KD. The authors' contribution lies in how the training signals are chosen, in the synthetic speech dialogues and in an efficient implementation of online distillation. This is a Preview version, and all figures come from the developers' own report.

For people building voice agents, the architecture matters: when speech understanding, reasoning and tool calling live in one open model, the system has fewer seams between components. For demonstration, the authors launched a platform called Apriel Live, where the model can be tried in live voice conversations.