Llama just got a dedicated transcription model.
Meta AI's model lineup never had a speech-transcription specialist — until now. Muse Voice Transcribe combines streaming recognition, speaker diarization, and endpointing into one model, giving teams a new option for self-hosting meeting notes and captions.
Why a dedicated transcription
model was overdue
Until now, Meta AI's model lineup was a three-way affair: Spark, Glimmer, and Llama 4. Muse Voice Transcribe, a dedicated real-time speech-transcription model, now joins that lineup — that's the news here. Its role is clearly separated from the existing models, which handle text generation and multimodal tasks; Llama's own writing and conversation abilities haven't changed.
The push to carve transcription out as its own model reflects real demand to run meeting notes and captions on self-hosted infrastructure. Many companies are wary of sending internal meeting audio to a cloud API, and the cost of calling a commercial transcription API for every meeting adds up as volume grows. Bundling streaming recognition, speaker diarization, and endpointing into one model is a design that answers self-hosting demand and privacy requirements head-on.
| Before (3-model lineup) | Now (4-model lineup) |
|---|---|
| Spark, Glimmer, Llama 4 only | + Muse Voice Transcribe |
| No dedicated transcription model | Purpose-built for real-time transcription |
| ASR, diarization, endpointing assumed separate | One model handles all three at once |
Three processes,
one model
ASR, speaker diarization, and endpointing run together, in one pass.
Muse Voice Transcribe combines streaming automatic speech recognition (ASR), speaker diarization, and endpointing into a single model. It's trained on more than 70 languages (with 25 described as extensively verified), and is said to handle mid-conversation language switching as well as conversations running over an hour. Details are published on Meta's model page.
It's available through three routes: the Meta AI Mac app, the developer coding tool Muse Code, and the Meta Model API. The Model API endpoint is compatible with the OpenAI SDK, so existing transcription code can reportedly be migrated by swapping only the base URL and model name.
How it stacks up against other
real-time transcription APIs
Based on Artificial Analysis's AA-WER Streaming benchmark.
Against major real-time transcription APIs, Muse Voice Transcribe's streaming WER (word error rate) is reported as the lowest of the group. That said, these figures come from reporting that includes Meta's own announcement — the AI Navigate editorial team has not independently re-measured them, and that caveat matters.
| Model | Streaming WER |
|---|---|
| Muse Voice Transcribe | 3.1% |
| Cartesia Ink-2 | 3.4% |
| ElevenLabs Scribe v2 Realtime | 3.6% |
| GPT Live Transcribe | 3.9% |
| Gemini 3.5 Transcribe Live | 4.0% |
Who benefits, and how
Engineers
The OpenAI-SDK-compatible endpoint means an existing transcription call can be trial-migrated by swapping only the base URL and model name — good news for teams wanting to keep self-hosting evaluation costs low.
Business & back office
Organizations wary of sending meeting audio to the cloud can run meeting notes and captions through a self-hosted pipeline. It's worth checking whether $3 per 1,000 minutes beats what you're already paying for transcription.
Marketers
With support for 70-plus languages and reported mid-conversation language switching, this opens up more use cases for real-time captions on multilingual webinars and live broadcasts.
Transcription no longer has to be a patchwork of separate tools.
What to do next
Don't just watch from the sidelines — run a small test first.
Compare the price first
Weigh the $3-per-1,000-minutes rate against your current cloud transcription API's monthly bill. Teams with more meetings and live streams will see a bigger gap.
Test small with the OpenAI-SDK compatibility
Before switching production traffic over, validate the setup by swapping only the base URL and model name.
Map out the self-hosting workload
Decide up front how much of GPU provisioning, availability, and keeping up with model updates your team will own versus what the cloud API used to handle.
Not all upside
The 3.1% WER figure rests on a third-party benchmark from Artificial Analysis plus Meta's own announcement — the AI Navigate editorial team has not verified it independently. Real-world accuracy will vary with microphone quality, background noise, and how much specialized vocabulary is involved.
The model is described as supporting "70-plus languages," but only 25 are said to be extensively verified; accuracy for the rest is an open question. How diarization behaves in meetings with more than 20 speakers also can't be judged from the public information alone. And to be clear, this addition only extends the speech-transcription side of things — Llama's text-generation and conversation abilities are unchanged. Teams that choose self-hosting also take on the availability and scaling work a cloud API used to handle on their behalf — a cost worth not overlooking.