Meta Launches Muse Voice Transcribe, Real-Time Speech Model Handling 20+ Speakers Across 70+ Languages
Meta Superintelligence Lab released Muse Voice Transcribe, its first real-time audio perception model, claiming state-of-the-art streaming speech-to-text with native speaker diarization. The model handles 20+ speakers and code-switching across languages, priced at $3 per 1,000 audio minutes.
Meta Superintelligence Lab (MSL) released Muse Voice Transcribe on September 1, 2026, a real-time audio perception model that Meta claims can distinguish between more than 20 speakers and switch between languages mid-sentence.
The model performs speaker diarization and endpointing natively in a single system — a departure from pipelines that typically chain separate models for these tasks together. According to Meta CEO Mark Zuckerberg, who demonstrated the model on X, Muse Voice Transcribe uses "adaptive delay to predict each token," waiting longer on ambiguous words and committing faster on straightforward ones to balance latency against accuracy.
Training and language coverage
Meta says the model was trained across more than 70 languages, with 25 validated at launch. It claims the system handles "code-switching" — sentences that blend vocabulary from multiple languages mid-utterance — and can sustain accuracy across hour-long sessions involving 20 or more distinct speakers. Meta has not published independent benchmark scores or a technical report alongside the release, so these performance claims remain unverified by third parties.
Availability and pricing
Muse Voice Transcribe is live now through three channels: it already powers dictation inside Meta's recently launched Meta AI Mac app, it's integrated into Muse Code (Meta's coding agent), and it's accessible to developers via Meta's Model API. Pricing is set at $3 per 1,000 audio minutes. A demo version is also available on Meta's research blog for anyone to test without an API key.
Meta AI chief scientist Alexandr Wang confirmed on X that the model is "live now via meta model API," already powering dictation across the desktop app and coding tool.
Competitive context
The release lands less than a week after Google shipped Gemini 3.5 Transcribe, an audio model with comparable multi-speaker and multilingual capabilities. Google has said it plans to build its model into Android and eventually Chrome. Meta has not disclosed similar plans to embed Muse Voice Transcribe into WhatsApp, Instagram, or its other consumer-facing products — for now, the model's visible reach is limited to the Mac app and developer API.
Muse Voice Transcribe is the latest in a rapid string of releases from MSL, which has shipped a dedicated coding agent, an open-weight language model, and the Meta AI Mac app all within the past few weeks.
What this means
Meta is positioning MSL as a high-velocity model shop, shipping specialized systems (audio, code, open-weight) in quick succession rather than a single flagship release cycle. The $3-per-1,000-minutes price point undercuts many enterprise transcription APIs, and native diarization removes a step most speech-to-text pipelines currently handle with bolt-on models — a real technical convenience if the claimed accuracy holds up under independent testing. But the near-simultaneous launch with Google's Gemini 3.5 Transcribe suggests real-time multilingual, multi-speaker transcription is becoming a contested commodity feature rather than a differentiator. The bigger question is distribution: Google is wiring its model into Android's operating system layer, while Meta's plans for its billion-plus-user apps (WhatsApp, Instagram) remain unstated. Without that integration, Muse Voice Transcribe risks staying a developer-tier product rather than a consumer-facing capability.
Related Articles
IBM Granite Switch 4.1 8B and 30B fold chat, decisions and speech-to-text into one endpoint
IBM has released Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, models that bundle a Decision adapter, a Guardian safety adapter and a speech-to-text model behind a single endpoint. IBM says the combined model uses roughly the same GPU VRAM as the base model alone. Benchmark claims are IBM's own, and context window and pricing were not disclosed.
Meta releases Muse Spark Open, an 800B-parameter open-weight model built for Apple M5 Ultra Macs
Meta released Muse Spark Open, an 800-billion-parameter open-weight model, on October 15, 2026. It is optimized for Apple's M5 Ultra Mac Studio and needs roughly 1TB of memory to run locally. Meta claims it outperforms GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.
Meta launches Muse Wonder agent model, gating it behind $20 and $100 monthly Muse app plans
Meta released Muse Wonder, a model it claims is its most advanced for powering digital assistants. Access requires a $20 or $100 monthly Muse app subscription, and a larger Muse Wonder Full variant is limited to Muse Code subscribers. API pricing has not been disclosed.
Elorian launches Glimpse, a visual reasoning model priced at $0.60/$3 per 1M tokens
Elorian, the Palo Alto startup that raised a $55M seed at a $300M valuation, has released Glimpse as a research preview with API access at $0.60 per 1M input tokens and $3 per 1M output tokens. The model can crop, pull video frames, and run code mid-reasoning. Benchmark claims come from Elorian's own evaluations.
Comments
Loading...