Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use

Apriel Voice 27B Preview is an open speech-language model built on Gemma 4 26B-A4B that uses a two-stage speech-adaptation recipe to unify speech understanding, reasoning, conversation, and tool calls while retaining text and image inputs.

A voice assistant must understand a spoken request, reason over context, sustain a conversation, and act through tools. Recognition accuracy is necessary but insufficient, and strong text performance does not establish reliable reasoning when the request arrives as audio. The paper targets an open speech-language model that retains text and image inputs and uses one backbone for responses and tool calls.

Apriel Voice 27B Preview starts from the post-trained Gemma 4 26B-A4B checkpoint and adds a speech interface. A frozen Whisper large-v3-turbo encoder feeds a learned projector with frame stacking and MLP layers. Stage one trains only the projector on multilingual transcription, translation, and continuation with the backbone frozen. Stage two updates the projector, language backbone, and mixture-of-experts router using speech tasks, synthetic speech-conditioned dialogues, tool interactions, and text replay. Source-dependent cross-entropy and forward KL supervision combine task targets with online teacher distributions from a frozen copy of the starting checkpoint. Training uses local teacher workers, one-batch lookahead, and fused CE/full-vocabulary FKL kernels; speech synthesis is external.

• In thinking mode, Apriel Voice scores 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and 65.51% task success on tau-Voice, with a mean latency of 1.91 seconds. §Abstract
• English EVA accuracy and experience pass@1 reach 51.27% and 74.13%. §Abstract
• Table 4 reports Apriel Voice average WER 12.61 and translation average BLEU 42.87. §9
• Table 19 reports tau-Voice Pass@1 and latency by domain: airline 69.33 and 2.22, retail 81.58 and 1.79, telecom 45.61 and 1.73. §E
• Table 18 reports EVA-A and EVA-X by domain: airline 59.3 and 66.7, ITSM 48.7 and 85.0, HR 45.8 and 70.7. §E
• Table 14 reports internal tau2-Bench average task success 70.95 in thinking mode and 49.14 in non-thinking mode. §C

• The paper states that the evaluations assess the combined method; component effects require controlled ablations.
• The paper states that conversion fidelity checks do not establish the correctness of a synthetic task or its target answer.
• The paper states that retention on other inputs remains an empirical question, especially when the teacher is incorrect or poorly suited to a new task.
• The paper states that full logits remain materialized; the memory-saving mechanism bounds temporary work rather than eliminating logit storage.

• Apriel Voice starts from the instruction-tuned Gemma 4 26B-A4B checkpoint and retains its text and image input modalities and output vocabulary. §3.1
• A frozen Whisper large-v3-turbo encoder feeds a projector with 47,722,496 parameters that maps 1280-dimensional features into 2816-dimensional decoder embeddings. §3.2
• Stage one trains the speech projector with the encoder and backbone fixed; stage two updates the projector, language backbone, and mixture-of-experts router. §3.3
• Stage one uses 18.17 M training examples over two completed epochs, each covering 32.2k hrs of audio. §4.2
• Stage two provides 52.6473 B total token exposure across three epochs, each containing 61.5k hrs of audio plus text data. §4.4
• Speech-conditioned dialogues use automatic span selection, one reference voice per conversation, and text retention for structured material. §5.2
• Each training source selects cross-entropy, forward KL, or a weighted combination; exposure weights determine sampling frequency separately from loss settings. §6.2
• Training uses local teacher workers, one-batch lookahead to overlap teacher inference with student computation, and fused CE/full-vocabulary FKL kernels. §7