Meta’s Superintelligence Labs just shipped its first real-time audio perception model, and the pricing is aggressive enough to make competitors sweat. Muse Voice Transcribe handles three jobs simultaneously — speech recognition, speaker identification for up to 20 people, and sentence boundary detection — without bolting on separate systems. It processes audio in 80-millisecond chunks and decides after each one whether to commit to a word or keep listening for more context. Easy words get output faster; harder ones get extra processing time. Meta trained this dynamic waiting behavior with reinforcement learning, optimizing for both low error rates and minimal delay.
Independent benchmarks from Artificial Analysis back up the claims. The model hits a 3.1 percent word error rate on English with just 0.16 seconds of latency after a speaker finishes. That edges out ElevenLabs Scribe v2 Realtime at 3.6 percent, AssemblyAI Universal-3.5 Pro Realtime at 4.0 percent, and Cartesia Ink-2 somewhere between 3.4 and 4.0 percent depending on configuration. OpenAI’s GPT-Realtime-Whisper is in the same race, and the company already slashed transcription prices in July to stay competitive. Meta’s response? $0.18 per hour of audio, undercutting Cartesia’s $4 and ElevenLabs’ $6.50 by an order of magnitude.
The model supports over 70 languages, with 25 tested in depth, and handles code-switching — speakers jumping between languages mid-sentence. Users can feed hints about language, keywords, or context to improve accuracy on proper nouns like ‘Meta’ or ‘Menlo Park.’ It’s available now in Meta AI and through the Meta Model API, powering voice dictation and Muse Code.
Meta frames this as the foundation for Zuckerberg’s ‘personal superintelligence’ vision — AI assistants that listen in on real conversations through smart glasses. That framing cuts both ways. Germany’s Federal Network Agency recently discussed banning Meta’s camera glasses before deciding against it, and privacy concerns around always-listening AI aren’t going away. The company hasn’t disclosed parameter count, training data volume, or audio sources, and isn’t releasing weights. That’s a notable shift for a lab that built its reputation on open models — and it follows the same playbook as Muse Spark 1.1 and 1.2: compete on price, not transparency.
