Paper review · November 3, 2026
Apriel Voice 27B Preview: An Open Model That Handles Speech, Reasoning and Tool Use in One Model
Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use
The authors attached a speech encoder and projector to Gemma 4 26B-A4B and applied two-stage training and On-policy Distillation to build an open speech-language model that reasons and calls tools from speech.
Problem
For a voice assistant to be useful, transcription alone is not enough. It must understand requests, reason about context, carry on a conversation and call tools to get the job done. Existing ASR→LLM→TTS cascades, or setups that hand work off to a separate dialogue model, split these functions across multiple components. Also, good text performance does not guarantee equally good reasoning on requests that come in as speech.
Method
Outputs from a frozen Whisper large-v3-turbo encoder are grouped in sets of 8 frames and mapped into the decoder embedding space by a projector of about 47.7 million parameters. In stage 1, the backbone is frozen and only the projector is trained on multilingual transcription, translation and speech continuation (18.17 million examples). In stage 2, the projector and the MoE backbone and router are updated together. Cross-entropy, Forward KL Distillation or a combination of the two is chosen per data source, and text replay preserves existing capabilities. A frozen copy of the original checkpoint serves as the Teacher and receives text with the same content. Training data is created by converting parts of internally synthesized text dialogues to speech with TTS. Training efficiency was improved by serving the Teacher locally, overlapping Teacher and Student computation, and fusing the full-vocabulary loss kernel.
Results
- In Thinking mode it scored 98.90 on Big Bench Audio and 81.13 on Speech-BFCL.
- Average task success rate on τ-Voice is 65.51%, with an average latency of 1.91 seconds.
- EVA-Bench English accuracy is 51.27%, and experience pass@1 is 74.13%.
- Stage 2 training ran on 8 nodes (16 Teacher GPUs, 48 Student GPUs) for about 3 days, covering a total of 52.6 billion tokens.
Limitations
Comparisons of the headline numbers with other models could not be found in the provided text. The authors claim it is strong relative to other open models. The dialogue training data was synthesized internally, and the authors also state that their fidelity checks do not guarantee the correctness of the synthetic tasks or the answers themselves. Speech output is left to an external synthesizer, and retention of visual capabilities still needs to be measured experimentally. It is still at the Preview stage.
Why it matters
As an open model that handles speech understanding, reasoning and tool calls with a single backbone, it offers an option for building voice agents such as call centers or internal help desks without a separate ASR or dialogue model. It is also a useful reference as a training recipe for adding a new modality to an existing LLM while keeping its original capabilities.
Key terms
- Speech projector
- A small neural network that maps acoustic features produced by a speech encoder into the language model's input embedding space.
- Forward KL Distillation
- A method that trains the Student model's next-token distribution to follow the Teacher model's distribution using KL divergence. Here it is used to preserve existing capabilities.
- Text replay
- A method that mixes in the original text data during adaptation to a new modality, preventing existing capabilities from being forgotten.
- MoE(Mixture of Experts)
- An architecture that activates only some Experts for each token. Of about 26 billion total parameters, this model uses only about 4 billion per token.
Original
Authors: Aman Tiwari, Akshay Kalkunte, Akintunde Oladipo, Jash Shah, Tara Bogavelli, Hoang Nguyen, Nishanth Madhusudhan, Vishal Juneja, et al. · Posted 2026-10-29 · License CC BY-NC-SA 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.