Jack Dorsey's Block announced Dessa, a voice model trained from scratch on human conversation with no text or LLM in the pipeline. The model is full-duplex (talks and listens simultaneously), runs locally on consumer hardware, and Block claims emotion is built into the model itself rather than driven by tags. There is no public access yet; Block is inviting safety researchers to test it.
Media coverage confirms Dessa runs locally on a single NVIDIA GPU and Block seeks safety researchers to test it before public access.
OpenAI announced expanded ChatGPT Voice capabilities on September 23, 2026. Voice can now use plugins (email, calendar, Slack), run on GPT-6 models Astra, Sol, and Luna, and operate in ChatGPT Work on web and mobile. The mobile rollout lets Plus and Pro subscribers trigger agentic workflows by voice, as voice-driven AI assistance competes with Anthropic's recent mobile handoff improvements.
OpenAI announced expanded ChatGPT Voice capabilities on September 23, 2026. Voice can now use plugins (email, calendar, Slack), run on GPT-6 models Astra, Sol, and Luna, and operate in ChatGPT Work on web and mobile. The mobile rollout lets Plus and Pro subscribers trigger agentic workflows by voice, as voice-driven AI assistance competes with Anthropic's recent mobile handoff improvements.
Sonive AI is an AI-powered speech and language platform that offers Text-to-Speech, Speech-to-Text, Translation, Audio-to-Audio Translation, and Voice Synthesis. It uses LLMs and speech models to enable multilingual communication and voice cloning for developers and content creators.
Sonive AI is an AI-powered speech and language platform that offers Text-to-Speech, Speech-to-Text, Translation, Audio-to-Audio Translation, and Voice Synthesis. It uses LLMs and speech models to enable multilingual communication and voice cloning for developers and content creators.
Developer 5uck1ess released Cicero, an open-source project enabling bidirectional voice conversations with AI agents. It builds on the author's tts-bench and blind voting arena for TTS models, and works with Hermes Agent and other harnesses.
Developer 5uck1ess released Cicero, an open-source project enabling bidirectional voice conversations with AI agents. It builds on the author's tts-bench and blind voting arena for TTS models, and works with Hermes Agent and other harnesses.
Quansys AI introduced Atlas-v1, a speech-to-speech voice model built in-house at its Noida team. The company claims the model matches frontier reasoning quality while answering in 0.32 seconds at a cost of roughly one cent per interaction. It matters to practitioners building low-latency, low-cost voice AI who currently rely on frontier-lab offerings.
Quansys AI introduced Atlas-v1, a speech-to-speech voice model built in-house at its Noida team. The company claims the model matches frontier reasoning quality while answering in 0.32 seconds at a cost of roughly one cent per interaction. It matters to practitioners building low-latency, low-cost voice AI who currently rely on frontier-lab offerings.
A developer released Red Curtain, a tool built with Claude that runs Fish Audio's Drama 3 text-to-speech model on a 16 GB AMD RX 9070 despite its official 48 GB NVIDIA GPU requirement. It keeps the FP8 experts (~32 GB) in system RAM and caches only used experts on the GPU. Performance is slower than realtime but workable for short takes, with a web UI for voice cloning and multi-speaker scripts.
A developer released Red Curtain, a tool built with Claude that runs Fish Audio's Drama 3 text-to-speech model on a 16 GB AMD RX 9070 despite its official 48 GB NVIDIA GPU requirement. It keeps the FP8 experts (~32 GB) in system RAM and caches only used experts on the GPU. Performance is slower than realtime but workable for short takes, with a web UI for voice cloning and multi-speaker scripts.