3.1% word error rate leads the streaming speech chart

Meta Superintelligence Labs released Muse Voice Transcribe on September 1. The streaming speech-to-text model scores a 3.1% word error rate on Artificial Analysis’s AA-WER Streaming benchmark, ahead of the live models from Cartesia, ElevenLabs, OpenAI and Google. Most of what’s new sits in how the model decides when to commit to a word, and that decision also sets its latency.

Meta calls it its first “real-time audio perception model,” The New Stack reports. Muse Spark, the first model of the Muse family, which Meta announced on April 8, could already turn speech into text but was never built for it. Since May 12 it has also run voice conversations in the Meta AI app, where users can interrupt, change topics or switch languages. Muse Voice Transcribe is the specialist built for live transcription.

The scoreboard, in English

On AA-WER Streaming the order is Muse Voice Transcribe at 3.1%, Cartesia Ink-2 at 3.4%, ElevenLabs’ Scribe v2 Real-time at 3.6%, GPT Live Transcribe at 3.9% and Gemini 3.5 Transcribe Live at 4%. Set against OpenAI’s live model, Meta’s figure works out to about a fifth fewer misrecognized words. Its lead over Cartesia is 0.3 points. The benchmark tests English speech only.

Telling speakers apart is where every real-time model still struggles. Meta’s model posts a 17.5% error rate on that task across several standard benchmarks, which again puts it first among the real-time systems compared.

The rest comes from Meta and has no independent number behind it. The company says the model separates more than 20 speakers and was trained on over 70 languages, 25 of them “extensively verified.” It also follows multilingual speakers who switch languages partway through a conversation and keeps going through conversations longer than an hour.

How the model decides when to write

Muse Voice Transcribe is an autoregressive multimodal model from the Muse Spark family: it produces output one token at a time, with each step conditioned on everything before it. Audio arrives in 80-millisecond chunks, 12.5 per second. Each chunk is compressed into a single soft token, a continuous vector fed to the model directly instead of an entry from its text vocabulary.

At every chunk the model chooses between two outputs. It can emit a text token, or it can emit a special “next audio” placeholder, which the system swaps for the next chunk of sound. Choosing the placeholder amounts to asking to hear more before writing anything. When the audio ends, an “empty audio” token tells the model nothing else is coming, and it flushes whatever text it has been holding back.

Because the model decides how much audio to take in before committing to a word, it also decides its own latency. Meta calls this “adaptive delay”: easy words come out almost at once, while difficult ones wait for more context. A fixed-window system has to pick one delay for every word in advance.

The model learns that tradeoff during reinforcement learning, where a reward for word error rate and a reward for delay are multiplied instead of added. With a product, a near-zero score on either side pulls the whole reward toward zero, so the model can’t make up for very slow output with accuracy, or for sloppy output with speed. Speaker detection runs on a similar mechanism.

Access, price and weights

The model is live in the Meta Model API, in Meta AI for Mac and in Muse Code, Meta’s terminal coding agent. API pricing is $3.00 per 1,000 audio minutes, or $0.18 per hour of audio.

The weights stay closed. A Meta spokesperson told The New Stack the company won’t release them. Meta did publish weights for Muse Glimmer, a 30-billion-parameter model it put on Hugging Face under Apache 2.0 in August. Anyone who wants Meta’s transcription on their own hardware has no option for now.

The model also enters a busy market. Per The New Stack, OpenAI, Google, xAI and Alibaba all shipped streaming transcription models within weeks of each other this summer, on top of the specialists already in the field, and the outlet expects a 0.3-point benchmark lead not to last long.

The New Stack also sees a reason Meta will keep working on it: every product the company really cares about, from its glasses to the Mac app, needs transcription to work as well as possible.