#语音

Introducing Grok Voice 3.0, the world’s most intelligent speech-to-speech model. It excels at resolving customer calls, responds in under a second, and costs less than other frontier models.

Introducing Sherpa: the most advanced fiction writing AI We accelerated from $250M in ARR to $500M because Sherpa helped increase content production by 1200% in 1 year Sherpa was trained on 5.5B hours of playtime with minute by minute dynamic retention data. 550K+ creators have produced 2.6M hours of content annualised using it Pocket FM is like Netflix for audio-only dramas, with our own pool of one-person studios. 10% of eligible writers on Pocket FM make >$200K One blockbuster produced >$100M in revenue 3 writers have become millionaires in <2 yrs We built Sherpa to enable anyone to make >$1M by writing world-class fiction stories: 1. The Idea: Drop a 1-2 sentence concept. Sherpa interrogates it like a veteran editor on tension, stakes, and psychology 2. World & Characters: It builds out the complete lore, tone, and character psychologies 3. Sub-Plot planning: Breaks the premise into arcs, arcs into episodes, and episodes into scenes 4. Scene-by-Scene Generation: Outlines and drafts entire episodes, with you able to steer, rewrite, or override anytime 5. Editorial Review: Stress-tests every draft for pacing, engagement drop-offs, prose, and coherence before it locks 6. One-Tap Production: Pick a voice, convert to audio drama, and publish directly to Pocket FM’s millions of listeners 7. Global Scale & Monetization: Revenue-share on performance, with automatic localization so you earn across international markets Test Sherpa for free here: https://pocketfm.com/sherpa _____________________________________________ Generic LLMs fail at serialized fiction because they lack a long-horizon narrative reward function. Sherpa solves this through three core technical leaps: 1. Narrative World Model (State Tracking & Retrieval): Context windows degrade over long runs. Sherpa constructs an evolving semantic knowledge graph tracking character states, secrets, and plot dependencies. High-speed retrieval surfaces exact context on demand, maintaining zero continuity decay across hundreds of episodes 2. Hierarchical Story Planner: When writing a 500-episode story like Naruto, you need to plan 100s of sub plots. Rather than generating linearly, Sherpa decomposes narrative across discrete levels: season -> arc -> sequence -> episode -> scene. Rather than generating everything upfront, like a generic LLM, Sherpa uses progressive planning and dynamic replanning. As the story evolves, it identifies what changed, traces the downstream impact, and replans only the affected parts. 3. Prose Engine (Trained on series' retention data): LLMs write robotically, but serial fiction needs emotion, tension, pacing, and dialogue that sounds like real people. Sherpa's Prose Engine was designed specifically for storytelling. It was built on 1B+ tokens of Pocket's own stories, trained by learning from what listeners engage with, where they drop off, and what keeps them hooked. Feedback is taken from specialized evaluator models that measure every scene against a 40-item checklist. (Evaluator models were benchmarked against human reviewers and matched them 80–90% of the time.) _______________________________________________ Owning distribution and creation puts us in a very unique spot. More shows -> More data -> Sherpa becomes better -> more creator success -> more creators -> more shows Pocket FM has already seen one $100M IP. I believe Sherpa will soon lead to dozens of single-person studios creating billion-dollar shows. Most people are scared of AI but I think it'll unlock more human creativity, help creators earn more, and bring the next great IPs to life. This will create millions of jobs and new income streams.

The new ChatGPT voice is quite impressive to use, really worth a minute to try it out on your phone. (while staying aware that the voice model is not going to be as smart as a full thinking model)

Duolingo is cooked 💀 GPT-Live fixes grammar while you speak

本地跑了一遍 VoxCPM2，最让我意外的不是它会说粤语、河南话，而是“声音”正在变成一种可以编辑的东西。 写一句提示词，就能指定年龄、音色、情绪和语速；再上传一段参考音频，它还能保留音色，重新控制表达方式。 以前语音模型追求的是“像不像本人”，现在已经开始解决“能不能按导演要求表演”。 当声音可以像图片滤镜一样被描述、复制和改写，它还算不算一个人的身份特征？ 视频里放了方言、情绪控制和真人声音克隆的实测效果。👇

Bro it’s June 2026. Stop hand editing your prompts. Hold down the dictation button and ramble for 10 minutes. Give the model every fragment, caveat, example, and vibe in your head. It is literally a large language model. If it’s superhuman at anything, it’s reconstructing latent intent from language.

People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way. We share our approach, early results, and a quick look at our model in action. https://thinkingmachines.ai/blog/interaction-models

Had tons of fun chatting with @sonyatweetybird about music and the future of entertainment

Announcing agentic performance benchmarking for Speech to Speech models on Artificial Analysis. We use 𝜏-Voice to measure tool calling and customer interaction voice agent capabilities in realistic customer service scenarios Even the strongest Speech to Speech (S2S) models today resolve only about half of realistic customer service scenarios end-to-end - a meaningful gap relative to frontier text-based agents on the same tasks. Voice channels introduce significant complexity: challenging accents, background noise, and packet loss, all while requiring fast responses, consistency across long multi-turn conversations, and reliable tool use. Performance also varies considerably by audio condition: in clean audio some models perform notably better, but realistic conditions continue to pose a challenge. Conversation duration also varies meaningfully across models, with implications for both customer experience and operational cost. About 𝜏-Voice: Our Agentic Performance benchmark is based on 𝜏-Voice (Ray, Dhandhania, Barres & Narasimhan, 2026), which extends 𝜏²-bench into the voice modality to evaluate S2S models on realistic customer service tasks. It measures multi-turn instruction following, support of a simulated customer through a complete interaction, and tool use against simulated customer service systems. The simulated user combines an LLM-driven decision model with realistic audio synthesis: diverse accents, background noise, and packet loss modelled on real network conditions. This complements our Big Bench Audio benchmark measuring intelligence and Conversational Dynamics (Full Duplex Bench subset) benchmark measuring conversational naturalness. Scores are the average of three independent pass@1 trials. We evaluate under realistic audio conditions using the 𝜏²-bench base task split across three domains: ➤ Airline (50 scenarios): e.g., changing a flight, rebooking under policy constraints ➤ Retail (114 scenarios): e.g., disputing a charge, processing a return ➤ Telecom (114 scenarios): e.g., resolving a billing issue, troubleshooting a service problem Task success is determined by deterministic checks against expected actions and final database state, consistent with the 𝜏²-bench evaluator. Key results: xAI's Grok Voice Think Fast 1.0 is the clear leader at 52.1%, averaging 5.6 minutes per conversation, the second-longest overall. OpenAI's GPT-Realtime-2 (High) (39.8%, 3.0 min) and GPT-Realtime-1.5 (38.8%, 4.8 min) follow, with Gemini 3.1 Flash Live Preview - High close behind at 37.7% (3.8 min). Speech to Speech is a fast evolving modality and we expect movement in rankings as we continue to add new models with these capabilities, and model robustness improves. Congratulations @xAI @elonmusk! See below for further detail ⬇️

Working as a public company in the ever-changing space of voice x AI: "The public market doesn't get it. You still need to deliver things quarter over quarter." Linda Sheng of @MiniMax_AI joins @EricNewcomer at Cerebral Valley Voice: