Moshi
PulseAugur coverage of Moshi — every cluster mentioning Moshi across labs, papers, and developer communities, ranked by signal.
6 day(s) with sentiment data
-
New InteractionBench benchmark evaluates real-time voice systems
A new benchmark called InteractionBench has been developed to evaluate real-time voice systems across multiple dimensions, including timing, task accuracy, whisper detection, and conversational grounding. The benchmark,…
-
Block unveils Dessa, a speech-only full-duplex AI model
Block's AI research lab has introduced Dessa, a novel full-duplex speech model trained exclusively on human conversation audio. Unlike many current voice agents that rely on text-first architectures, Dessa operates dire…
-
New attack impersonates voices using leaked speech representations
Researchers have developed a novel attack called \"stia\" that can impersonate a speaker's voice using only a short leaked speech representation, without needing the original audio or its transcript. This method convert…
-
Moshi CEO Anekha Sokhal on AI for Market Data and Trader Workflows
Anekha Sokhal, CEO of Moshi, discussed her journey from energy derivatives trading at JPMorgan to founding an AI-powered market data recommendation engine. Moshi aims to integrate market data, workflow interfaces, and t…
-
Podcast Explores Recommendation Engines for Commodity Traders
This podcast episode features Anekha Sokhal and Moshi discussing recommendation engines, particularly their application within commodity trading. The conversation likely delves into the technical aspects and practical u…
-
Speech models evaluated on natural conversation flow and recall
A new arXiv paper evaluates the conversational abilities of speech models, focusing on their decision-making at conversational pauses. The study found that hosted APIs like GPT Realtime and Gemini 3.1 Flash Live tend to…
-
New research explores full-duplex dialogue models, memory, and turn-taking
Researchers are developing new methods to improve full-duplex spoken dialogue models, which allow for simultaneous listening and speaking. One approach, PERSIST, introduces a memory system that tracks who said what and …
-
New method suppresses spurious speech in full-duplex LLMs
Researchers have identified and addressed an issue in full-duplex speech LLMs where models like Moshi and PersonaPlex inappropriately initiate speech during prolonged user silence. The problem stems from a sudden spike …
-
RetroThinker framework boosts SpeechLLM reasoning accuracy
Researchers have developed RetroThinker, a novel post-training framework designed to enhance the reasoning capabilities of speech-based large language models (SpeechLLMs). This framework enables models like Moshi to sel…
-
New AV-STE system enhances dialogue models with audio-visual speech token restoration
Researchers have developed AV-STE, a novel modular front-end system designed to enhance audio-visual speech token processing for spoken dialogue models. This system aims to improve the robustness of full-duplex dialogue…
-
New datasets and models advance full-duplex spoken dialogue research
Researchers have introduced new datasets and models focused on full-duplex spoken dialogue systems, which enable more natural, real-time conversational interactions. The DuplexDrama dataset offers over 2,000 hours of sy…
-
Mimi codec's semantic tokens linked to phonetic realizations
Researchers have investigated the Mimi codec, a component of the Moshi language model, focusing on its 2048-token semantic codebook. Their findings indicate that standard ABX experiments are insufficient for understandi…
-
Metronome system bounds AI model cache for real-time interaction stability
Researchers have developed a new system called Metronome designed to improve the real-time serving of interactive AI models. These models, such as Moshi, MiniCPM-o, and Qwen Omni, face a critical issue where sustained l…
-
New dialogue system integrates real-time facial generation with speech
Researchers have developed Moshi-Face, a novel full-duplex spoken dialogue system that integrates facial generation with audio processing. This system utilizes a VQ-VAE to encode facial data into discrete tokens and a F…
-
BayLing-Duplex enables native full-duplex speech dialogue with single LLM
Researchers have developed BayLing-Duplex, a novel full-duplex speech language model that enables simultaneous listening and speaking without relying on external turn-taking modules. This single autoregressive LLM can m…
-
Moshi dialogue models show synchronized internal states and predict turn-taking
Researchers have explored how full-duplex speech dialogue models coordinate their internal representations during interaction. By simulating dialogues between two instances of the Moshi model, they observed strong repre…
-
Thinking Machines previews interaction models for real-time AI collaboration
Thinking Machines has introduced a research preview of interaction models designed for native, real-time collaboration. These models process audio, video, and text simultaneously, allowing for continuous thought, respon…
-
New methods boost full-duplex speech models for better interaction
Researchers have developed new methods to enhance full-duplex speech models, enabling more natural and interactive conversations. One approach focuses on improving interactivity axes like pause handling and turn-taking …
-
Sakana AI's KAME architecture injects LLM knowledge into speech AI without latency
Sakana AI has developed KAME, a novel tandem architecture for speech-to-speech AI that aims to combine the speed of direct systems with the knowledge depth of LLM-based approaches. KAME operates with two asynchronous co…
-
Josh Talks launches first full-duplex Hindi conversational AI model
Researchers have developed the first open and reproducible full-duplex spoken dialogue system for the Hindi language. This system, named Human-1, adapts the Moshi architecture and was trained on over 26,000 hours of rea…