Meta Voice Model Tops Leaderboard with 3.1% Word Error Rate, Priced at One-Fifth of Google

On September 1, 2026, Meta Superintelligence Labs officially released Muse Voice Transcribe, integrating streaming speech recognition, speaker diarization, and endpointing into a single real-time model. According to independent evaluations by Artificial Analysis, the model ranked first on the AA-WER Streaming speech transcription leaderboard with a 3.1% word error rate and also took the top spot in public speaker recognition benchmarks. API pricing is set at $3 per thousand audio minutes (equivalent to $0.18 per hour), approximately one-fifth of Google Cloud Speech-to-Text's standard price.

This is the first complete public scorecard for a speech AI product since Meta Superintelligence Labs was established. It pressures competitors on both accuracy and price simultaneously. Muse Voice Transcribe is not a simple engineering optimization, but a redesign at the architectural level.

Single-Model Architecture: One Brick Replaces Three Pipelines

Traditional commercial speech recognition systems typically handle three functions separately: automatic speech recognition (ASR) converts sound into text; speaker diarization determines "who is speaking"; endpointing determines "whether the speaker has finished." These three modules are independent of one another. Their outputs need to be stitched together, latency accumulates across the pipeline, and errors propagate between modules.

According to Meta's official blog, Muse Voice Transcribe fuses these three capabilities into a single autoregressive model that processes audio in real time in 80-millisecond frame units, with no post-processing steps required. The model introduces a reinforcement-learning-driven "adaptive latency" mechanism — it dynamically adjusts output timing based on the recognition difficulty of the current vocabulary, rather than pushing out text on a fixed beat. When a word is sufficiently certain, the model outputs it immediately; when context is ambiguous, it waits a few dozen extra milliseconds in exchange for higher accuracy.

According to Artificial Analysis data, Muse Voice Transcribe's final transcription achieves a word error rate of 3.1% and is output 0.16 seconds after speech ends; Cartesia Ink-2 has a word error rate of 3.4% but a latency of 0.43 seconds; ElevenLabs Scribe v2 has a word error rate of 3.6% with a latency of 0.14 seconds; Google Gemini 3.5 Transcribe Live has a word error rate of 4%. Muse has found the optimal balance between latency and accuracy.

The speaker diarization capability supports more than 20 people present at once and natively handles multilingual input. Officials say the model has been trained on more than 70 languages, of which 25 have undergone large-scale validation, and it supports seamless language switching within a single sentence (code-switching). The demo scenario in the official blog features eight people in the same room mixing English and Chinese; the model can distinguish each speaker's content word by word.

Pricing Strategy: Entering the Developer Market at One-Fifth of Google's Price

According to Meta's developer documentation, Muse Voice Transcribe is served through the Meta Model API at $3 per thousand audio minutes, equivalent to $0.18 per hour. Compared with Google Cloud Speech-to-Text's standard pricing of roughly $0.96 per hour, Meta's price is approximately one-fifth.

Looking across the market, Cartesia Ink-2 is priced at $4 per thousand minutes, while ElevenLabs Scribe v2 and Deepgram Flux are both priced at $6.5 per thousand minutes. Among the four major streaming speech products, Muse Voice Transcribe has the lowest price and the best word error rate.

This pricing logic is consistent with Meta's strategy in the large language model space in recent years: leveraging infrastructure scale advantages to lower unit costs, opening up the developer market with price, and then consolidating market share through ecosystem stickiness. The historical backdrop of the Llama open-source model series pulling down the overall pricing of proprietary LLM APIs makes this speech pricing move unsurprising.

Competitive Landscape Reshaped: Google Is Not the Only One Under Pressure

Google is under pressure on both accuracy and price, and its Gemini 3.5 Transcribe Live's 4% word error rate was the worst result in this round of Artificial Analysis evaluations. Google Cloud's market advantage with enterprise customers comes more from ecosystem integration (connections with products such as BigQuery and Workspace). This moat will not be affected in the short term, but in pure API selection scenarios, a fivefold cost gap demands a more compelling justification.

ElevenLabs' position is more delicate. Its Scribe v2 Realtime is close to Muse in latency (0.14 seconds), but its word error rate (3.6%) is higher, and its price ($6.5 per thousand minutes) is more than double Muse's. ElevenLabs' core strength lies in speech synthesis (TTS) rather than transcription (STT). Muse Voice Transcribe is directly attacking ElevenLabs' emerging direction of multimodal voice capability expansion.

For small and medium-sized developers and startups, this is an unequivocal positive. The previous cost barrier of real-time speech transcription forced many application scenarios requiring high-precision multilingual recognition — meeting minutes, real-time subtitles, cross-language customer service — to choose offline batch processing or accept higher operational costs. Muse now provides real-time streaming capability at equal or even lower cost.

The model is currently offered only in API form, without open model weights. This means privacy-sensitive scenarios (audio transcription in finance, healthcare, and legal industries) or enterprise customers with strong on-premises deployment requirements cannot deploy it in a private environment. In this dimension, the open-source Whisper series remains an irreplaceable option, although it falls short of Muse in both streaming real-time capability and speaker diarization.

On-Device Integration: The Mac Platform Is the Testing Ground

Alongside the API release, Meta has integrated Muse Voice Transcribe into the Meta AI for Mac client and Muse Code. According to 9to5Mac, the Mac integration allows users to invoke real-time voice dictation in any application, meaning it works without relying on native support from a specific app.

The Mac platform has long seen competition between Apple Dictation and third-party tools, but real-time tools supporting multiple speakers, mixed languages, and low latency are almost nonexistent. By embedding Muse directly into system-level Mac scenarios, Meta gains both a product demonstration and a channel for collecting real user behavior data. If the Mac client performs well, the next natural landing point is AR glasses — Meta has always regarded smart glasses as the core terminal for personal AI assistants, and "distinguishing what each person says in a noisy real-world environment" is precisely the hardest perception problem in the glasses scenario.

Strategic Assessment: How Wide Is the Accuracy Leadership Window

The following is analysis and judgment, not confirmed fact.

Muse Voice Transcribe's current lead is real, but the window is limited. Improving speech recognition accuracy is a relatively well-defined path in engineering — more data, longer training time, finer domain adaptation — and all players are advancing along it simultaneously. Artificial Analysis leaderboard figures fluctuate frequently as models are updated; the 3.1% versus 3.4% gap could be erased at any time in the next refresh.

What Meta truly makes difficult to replicate is its structural advantage in pricing. Behind the $3 per thousand minutes is the scale amortization of Meta's self-built compute clusters — a cost gap that cannot be closed through technical iteration alone. If this pricing strategy persists, the overall pricing center of gravity in the speech API market will be pulled downward, which means sustained profit pressure for Google, ElevenLabs, and Deepgram.

What has been disclosed about the model so far is mainly technical metrics; it lacks validation data on long-term stability in large-scale production environments. That is the biggest current uncertainty. Whether Meta will open Muse's weights at some point will directly determine whether it can enter the privacy-sensitive enterprise market — precisely the most lucrative segment for speech recognition monetization.

For developers making selection decisions today, there is a fairly clear framework: if the use case requires real-time streaming, multiple mixed speakers, multilingual support, and data can be processed via API, Muse Voice Transcribe is currently the optimal choice on both cost and accuracy; if the use case involves sensitive data or requires local deployment, the open-source Whisper series remains the safer path, but the engineering problems of streaming and speaker diarization must be solved in-house. Between these two paths, no solution currently exists that is both open-source and able to compete head-on with Muse.