Meta Muse Voice Transcribe hits 3.1% WER, ships in the API today
Best-in-class real-time transcription with diarization for 20+ speakers and 70+ languages, at $0.18/hr. Plus: World Labs' Atlas and Google Pics ship.
Copy markdown
One model does transcription, diarization, and endpointing
Muse Voice Transcribe is Meta Superintelligence Labs' first real-time audio perception model, handling streaming ASR, speaker diarization, and endpointing in a single pass — no separate pipeline to stitch together. It's live now on the Meta Model API at $3 per 1,000 audio minutes ($0.18/hour).
It's the new streaming-accuracy leader
3.1% word error rate on the English AA-WER streaming benchmark — ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 Real-time (3.6%), OpenAI's GPT Live Transcribe (3.9%), and Gemini 3.5 Transcribe Live (4%). Diarization error is 17.5%, distinguishing 20+ speakers with no post-processing.
~0.16s to final text via 'adaptive delay'
The model learns per word how much audio context it needs before committing, trained with RL that multiplies the accuracy and latency rewards instead of adding them. The payoff: final transcripts in about 0.16 seconds — fast enough for live captioning, voice agents, and meeting notes.
70+ languages, code-switching, hour-long calls
Trained on 70+ languages (25 verified at launch, including Chinese, French, Hindi, Japanese, Spanish, and Vietnamese), it handles within-sentence language switching and conversations running over an hour. The catch: no open weights — you reach it through the Meta Model API, Meta AI for Mac, or the Muse Code IDE.
Elsewhere: World Labs' Atlas world model — but hands off for now
Fei-Fei Li's World Labs unveiled Atlas, a multimodal model that generates up to a minute of 1440p video with pixel-perfect camera control and reconstructs 3D scenes (point clouds, Gaussian splats) from just 1-3 images. It's early access for select partners only — no public API, weights, or pricing yet.
Elsewhere: Google Pics takes on Canva
Google shipped Pics, a Workspace-native AI image creator and editor built on Nano Banana 2, to AI Pro and Ultra subscribers. It does targeted object edits, in-image text translation, and 4K upscaling — aimed squarely at precision editing rather than one-shot generation.