Moshi (AI model): news timeline & CVEs

HiPLEX: Hierarchical Policy Factorization for Full Duplex Speech Language Models

HiPLEX factorizes full-duplex speech policies so reinforcement learning can improve timing and content separately.

HiPLEX is a reinforcement-learning framework that factorizes a pretrained full-duplex text policy into a control policy choosing pad, epad, or con, and a conditional content policy that emits a token only when con is selected. Timing advantages use event-causal masks from generated speech, while an LLM-judge semantic advantage trains the content factor through the existing text head. Across three Moshi seeds on Full-Duplex-Bench v1, it reduces takeovers during pauses and backchannels and shortens post-interruption latency versus GRPO with comparable judged quality. On Moshi and PersonaPlex it matches pooled human turn-timing and backchannel-rate marginals better than GRPO.

Hugging Face daily papersOct 5, 20:00 UTC · 3d agoAI research1

Tokens Change, Structure Endures: Spectral Watermarking for Generated Speech

Redwing keeps generated-speech watermarks intact through retokenization using a spectral substitution basis.

Researchers propose Redwing, a training-free token-level watermark for generated speech that models retokenization substitutions as a graph and embeds the mark in a Laplacian spectral basis. On the Moshi full-duplex system, eight consecutive Mimi resynthesis passes yield 80.7% true-positive rate at a 1% false-positive rate, versus 8.3% for KGW and at most 7.3% for WMAR. Across three other neural codecs, eight-pass TPR is 77.5-93.0%, and the gains transfer to TTS models at a speech-quality cost close to KGW.

arXiv cs.CRSep 27, 17:13 UTC · 11d agoAI research

RetroThinker: Enabling Retrospective Thinking in Speech LLMs

RetroThinker is a post-training framework letting the Moshi speech LLM self-correct reasoning mid-stream, adding 11% GSM8K accuracy at similar latency.

Researchers introduce RetroThinker, a multi-stage post-training framework that equips the Moshi speech LLM to verify and forward-correct chain-of-thought steps during streaming inference. It combines supervised fine-tuning on curated retrospective thinking data with length-based direct preference optimization (DPO). On GSM8K it achieves an 11% absolute accuracy gain over non-retrospective baselines at comparable latency.

arXiv cs.AI / cs.LG / cs.CLSep 10, 17:41 UTC · 28d agoAI research1

Moshi

Timeline3 articles

Appears with

HiPLEX: Hierarchical Policy Factorization for Full Duplex Speech Language Models

Tokens Change, Structure Endures: Spectral Watermarking for Generated Speech

RetroThinker: Enabling Retrospective Thinking in Speech LLMs