Fish Audio just dropped Drama 3, an open-source 35B MoE speech model built for complex audio storytelling. It can juggle up to 16 distinct character voices in a single script, add contextual tags like sighs or laughs, and patch incorrect words without re-generating the entire audio track. It's basically a full cast in a single model weights file.
Cactus Compute just released Whistle, a tiny speech recognition model that fits into a single 16.9 MB file. It runs entirely on the CPU with zero dependencies, delivers transcription and word-level timestamps across seven languages, and beats the 145 MB Whisper base model on several standard benchmarks.
Suno just launched Speech into open beta, an audio model that generates spoken voice and original background music together in a single track. Users simply type a poem, idea, or script, describe the voice and musical backdrop, and get back a finished audio piece. Why it matters: it collapses voice generation and scoring into a single prompt.
Google just rolled out Guided Vision in Gemini Live on Android, turning the AI assistant into a real-time conversational visual partner for blind and low-vision users. Built with accessibility org Aira, it lets users share their camera to hear surroundings described, read small text, and locate dropped items—with Gemini even offering verbal reframing cues if your angle is off. Multimodal AI is finally finding its most meaningful real-world utility.
Microsoft just launched MAI-Voice-2.1 and MAI-Voice-2.1-Flash, two text-to-speech models capable of instant zero-shot voice cloning. The Flash variant drops inference latency to just 45ms at $15 per million characters, designed to eliminate awkward processing pauses in real-time customer support agents across 23 languages.
Tavus just unveiled Griffin, a full-duplex human interaction model that holds real-time video calls with lifelike expressions and gestures. Unlike turn-based bots, it watches and listens simultaneously, nodding mid-sentence or reacting to pauses, tricking 48% of live study participants into thinking it was human. Griffin-Lite is currently available to early testers.
Inception Labs just rolled out Mercury Voice for enterprise customers, a diffusion LLM tuned specifically for conversational agents. It hits a median time to first answer token of 320 milliseconds—5.9x faster than GPT-6 Luna—without sacrificing tool-calling and reasoning performance. It drops straight into OpenAI-compatible voice pipelines, effectively killing the awkward pause in phone agents.
ElevenLabs just released Eleven v4 alongside Eleven v4 Turbo, its most expressive text-to-speech models yet. The update lets creators direct emotion and sound effects using inline tags, cuts voice cloning audio requirements to just 10 seconds, and supports over 90 languages. With median inference latency clocked at roughly 100ms, Turbo enables natural, expressive voice agents that respond without awkward conversational pauses.
Fermion Research just open-sourced Phonon-2, a lightweight English speech recognition model under 164MB. It transcribes an hour of audio in roughly 20 seconds on an M5 MacBook Air while scoring a 5.21% average word error rate. It proves high-accuracy ASR no longer needs multi-gigabyte models or cloud dependencies.
ElevenLabs just launched a student tier offering college students in the US, Canada, the UK, Australia, and the EU three months of free access to its creative suite, agents platform, and APIs, alongside a free year of ElevenReader Ultra. The move aims to put professional-grade voice AI into the hands of young builders and researchers, with broader international rollout planned next.
Google just launched Gemini 3.8 Live with Live Avatar, bringing real-time visual personas to its conversational AI in Gemini Enterprise. The system pairs near real-time video generation with speech, delivering fluid turn-taking and accurate lip-syncing across 97 languages. It can even call background tools without interrupting active dialogue.
OpenAI just rolled out a major ChatGPT Voice upgrade worldwide, letting users connect tools like email, calendar, and Slack while running on new GPT-6 models. It also integrates into ChatGPT Work across web and mobile, enabling users to build docs, decks, sites, and spreadsheets purely by speaking.
Researchers just unveiled ACTx486, an experimental system that transforms pre-recorded videos into interactive media you can actually talk to. Viewers can interrupt the host with questions, request visual diagrams, or alter scene details before the clip seamlessly resumes. Full interactions still take minutes to generate, so the current demo is pre-rendered while the team aims for sub-six-second latency.
Google just launched Gemini 3.8 Flash TTS and Flash-Lite TTS, bringing prompt-based voice creation across more than 100 languages and dialects. Creators can design entirely new vocal personas from scratch or clone voices from a 30-second audio sample, with line-by-line directorial control over pacing, accents, and emotional interjections. Both models are available now in the Gemini API and Google AI Studio.
YouTube just unveiled Ask Music, an in-app conversational feature that lets listeners explore its massive catalog, customize queues, and dig into artist lore via chat. Subscribers also get a personalized AI audio guide that delivers weekly spoken previews of recommended podcasts hands-free. It basically turns music streaming into an interactive, voice-driven radio tailored to your taste.
ElevenLabs just launched Studio 4.0, turning its editor into an all-in-one AI production suite. Creators can now generate video, images, sound effects, music, and voiceovers across 32+ languages directly on the timeline without switching tools. It also adds Studio Agent, an AI co-editor that drafts rough cuts and syncs audio frame by frame.
ElevenLabs just introduced Studio 4.0, an all-in-one AI audio editor designed for podcasters and video creators. The platform features over 10,000 voices, background music generation, and instant speech correction that regenerates cloned audio simply by editing the script. It also adds Studio Agent, a co-editor that drafts scripts and arranges sound effects from a single prompt.
Edge0 just released Audio8-ASR-Infinite, a bilingual streaming speech recognition model that decodes up to 12.5 times a second. By pairing a rolling KV cache with an adapted vLLM setup, it keeps latency and memory usage strictly constant even during nonstop 24/7 audio streams.
Alibaba just released Qwen3.8-LiveTranslate, a real-time interpretation model that cuts average latency to 2.3 seconds across 60 languages. It reconstructs speech into an interleaved audio-text stream and clones speakers' voices so translations sound just like them, making everyday bilingual meetings far less clunky.
xAI just released Grok Voice Transcribe 2.0, delivering twice the transcription accuracy of its predecessor at the same rate of $0.10 per batch hour and $0.20 for streaming. It ranked first among 32 streaming models on Artificial Analysis, slashing multilingual short-phrase error rates from 20.6% down to 6.8%. Atlassian Loom has already adopted it across its entire platform.
