Meta Muse Voice Transcribe Reality Check: 3.1% Streaming WER, 0.16s Latency and English-Benchmark Limits

Meta introduced Muse Voice Transcribe on September 1, 2026 as the first real-time audio-perception model from Meta Superintelligence Labs. The model combines streaming automatic speech recognition (ASR), speaker diarization and speech endpointing in one autoregressive system. Meta says it is trained across more than 70 languages, with 25 extensively verified for the initial release, and that it supports code-switching, language/keyword/context biasing, hour-plus audio and more than 20 speakers.

The launch is genuinely interesting for voice-agent and meeting-transcription developers, but several headline numbers need labels. The strongest independent speech-to-text result we found is 3.1% final-transcript word error rate (WER) at about 0.16 seconds after detected end of speech on Artificial Analysis's AA-WER Streaming benchmark. That is a strong English streaming result under a specific API benchmark harness. It is not evidence that the same 3.1% error rate holds across all 70+ training languages, every accent, every noisy room or every multi-speaker conversation.

Independent streaming result: 3.1% WER at about 0.16 seconds

Artificial Analysis reported on September 1 that Muse Voice Transcribe took the top final-transcript accuracy position on its streaming leaderboard at 3.1% WER and 0.16 seconds to final transcript after end of speech. Contemporary benchmark reporting places Cartesia Ink-2 with semantic endpointing at about 3.4% WER and 0.43 seconds, and ElevenLabs Scribe v2 Realtime at about 3.6% and 0.14 seconds.

Those numbers should be read as paired accuracy-and-latency coordinates, not as a single ranking. A model can improve final accuracy by waiting longer. Artificial Analysis explicitly evaluates streaming ASR on both WER and time after end of speech, and its methodology includes network delay. That means a developer choosing a voice stack should compare points near the latency budget they actually need, not simply choose the lowest WER.

Relative to a 4.0% WER result, 3.1% is a 0.9 percentage-point absolute reduction and a 22.5% relative reduction in word errors. But that arithmetic does not make every 4.0% system directly inferior: endpointing configuration, model mode, provider path, network conditions and latency target can all differ.

The benchmark is roughly eight hours and is English-focused

Artificial Analysis's current speech-to-text methodology says it independently tests model-provider API endpoints on roughly eight hours of audio. The AA-WER result is weighted 50% AA-AgentTalk, 25% VoxPopuli-Cleaned-AA and 25% Earnings22-Cleaned-AA.

The public VoxPopuli subset is explicitly English. Earnings22 is an English corporate-earnings dataset, and AA-AgentTalk is designed around voice-agent speech. The benchmark therefore provides useful evidence about English streaming transcription, including conversational and technical speech, but it should not be presented as a 70-language multilingual score.

The methodology also matters for the 0.16-second latency figure. Artificial Analysis starts its streaming latency clock after Silero VAD detects end of speech. For systems that support forced endpointing, the harness sends an endpoint signal and times the next final transcript; for systems without that path, it uses documented fallback logic. WER is computed against verified reference transcripts after normalization. This is a reproducible operational definition, but it is not the same as "the model always finishes 160 milliseconds after a human stops talking" in every application.

Streaming and batch transcription are different benchmark families

Muse's 3.1% result is a streaming result. It should not be ranked directly against the best batch/offline speech-to-text numbers, where a model can process a complete recording with future context and different latency assumptions.

For example, Artificial Analysis has reported lower WER values for some batch systems. Those results answer a different question: how accurately and quickly a complete file can be transcribed after it is available. Voice agents care about a live accuracy-latency frontier; archive transcription often cares more about final accuracy and throughput.

A fair comparison therefore keeps streaming and batch leaderboards separate, just as OP08 keeps SWE-bench Verified and SWE-bench Pro separate in coding coverage.

Why "adaptive delay" is the core architectural idea

Meta describes Muse Voice Transcribe as an autoregressive multimodal model in the Muse Spark family. Audio is processed in 80 millisecond chunks, or 12.5 chunks per second, with each chunk converted into a soft token. At each step the model can either continue listening or emit text.

That gives the model direct control over how much future acoustic context it waits for before committing a word. Meta calls this adaptive delay. The company says reinforcement learning combines a WER reward and a delay reward so the model can wait longer for difficult words while committing easier words sooner.

This is a more meaningful architectural claim than simply calling the system "low latency." In streaming ASR, fixed delay creates a blunt tradeoff: wait more and accuracy often improves, but conversation feels slower. A learned per-word policy can potentially spend latency where it is most useful.

The launch benchmark supports the claim that Muse is competitive on the accuracy-latency frontier, but independent replication across different endpointing policies, networks, accents and noise conditions would be more informative than a single leaderboard coordinate.

Diarization: 17.5% is promising, but it is still an error rate

Meta also reports an average diarization error rate (DER) of 17.5% across AMI-IHM, AMI-SDM and VoxConverse in its public benchmark comparison, lower than the comparison systems shown in the launch material. Meta says the model can distinguish more than 20 speakers and can process audio longer than one hour without a required external post-processing stage.

Two caveats matter.

First, this 17.5% result is presented through Meta's launch evaluation over public datasets. We did not find an independent reproduction matching the same model version, streaming settings, speaker conditions and scoring protocol. It should therefore be labeled a Meta-reported public-benchmark result, not silently upgraded to an independently reproduced measurement.

Second, DER is not "17.5% of speakers are wrong" or "17.5% of words have the wrong name." Diarization metrics score how accurately speech time is assigned to speakers, including attribution and segmentation errors under a defined protocol. A 17.5% DER is good evidence that real-time speaker labeling is improving, but it is not a reason to treat speaker identity as infallible in legal, medical, compliance or meeting-record systems.

The launch comparison also mixes streaming and offline competitors in some rows. Offline diarization can use the whole recording before assigning speakers, so mode and post-processing must be preserved when comparing systems.

70+ trained languages is not the same as 70+ independently benchmarked languages

Meta says Muse Voice Transcribe was trained on 70+ languages and that 25 were extensively verified for the initial release. It demonstrates multilingual transcription and Mandarin-English code-switching, and it supports language, keyword and context biasing.

That is meaningful product coverage, especially for bilingual speech where a conversation can switch languages within one sentence. But the independent AA-WER Streaming evidence reviewed here is English-focused. We did not find a comparably broad independent multilingual evaluation covering the 25 validated languages, much less the full 70+ training set.

A launch-week X user, @supbagholder, reported that Muse handled their English/French "Franglais" unusually well compared with their existing workflow. That is useful first-hand developer feedback, but it is one self-selected anecdote. It cannot establish multilingual average accuracy or prove superiority across French-English code-switching.

The correct evidence hierarchy is therefore:

• Meta verifies training breadth, 25 launch-validated languages and code-switching support.
• Artificial Analysis provides strong independent English streaming evidence.
• Early user reports can identify promising cases and failure modes, but they are not a representative benchmark.

Public feedback includes both enthusiasm and benchmark criticism

Launch-week public discussion is not uniformly celebratory. Alongside users praising speed and code-switching, diarization researcher Hervé Bredin (@hbredin) publicly questioned a benchmark comparison that omitted pyannoteAI, writing in an X discussion that the absence was worth noticing. Techmeme archived that disagreement on September 2.

That criticism does not invalidate Meta's result. It highlights a common benchmark-selection issue: a chart can be technically correct for the systems included while still being incomplete for a practical buyer deciding among available diarization stacks.

No broad "developer consensus" should be inferred from a handful of X posts. The useful next step is an open, reproducible evaluation that includes strong specialist diarization systems, fixes streaming/offline mode, reports overlap handling and speaker-count distributions, and publishes the scoring configuration.

Pricing and access

Meta's research page confirms that Muse Voice Transcribe is available through the Meta Model API, Meta AI for Mac and Muse Code. Launch-day coverage reports API pricing of $3 per 1,000 audio minutes, equivalent to $0.18 per audio hour.

That price is attractive for large transcription volumes, but it should be checked against the current Meta developer pricing page before committing production spend. The direct developer documentation endpoint was rate-limited during this verification pass, so this article treats the $3/1,000-minute figure as launch pricing corroborated by multiple contemporaneous reports rather than pretending it was re-read from a live billing table in this run.

Mark Zuckerberg also said the hosted Meta Model API has a zero-data-retention tier. That is an API data-handling option, not proof that every client integration is on-device or that no metadata is retained under every plan. Production users should verify the current terms and enterprise data-processing details for their account.

Meta's launch material does not announce an open-weight release for Muse Voice Transcribe. The verified access paths in this review are hosted Meta services.

Context, latency and model-size information that is still missing

Unlike text LLM releases, a real-time speech model does not have a directly comparable token-context figure. Meta documents support for audio exceeding one hour, but we did not find a public parameter count, model memory footprint, downloadable checkpoint, hardware requirement for self-hosting, or a fixed maximum audio duration beyond the demonstrated hour-plus capability.

The 0.16-second Artificial Analysis number is end-of-speech-to-final latency under its benchmark protocol; it is not full mouth-to-agent response latency. A voice application still adds microphone buffering, endpoint detection, network transport, downstream language-model reasoning, tools and text-to-speech.

Developers should therefore measure complete turn latency in their own stack, including p50/p95 and failures, instead of treating ASR finalization time as total conversational latency.

SWE-bench Verified and SWE-bench Pro are not applicable scores

Muse Voice Transcribe is a speech-perception model. We found no meaningful SWE-bench Verified or SWE-bench Pro score for it, and none should be invented or borrowed from the Muse Spark text-model family.

SWE-bench Verified and SWE-bench Pro measure software-engineering agents under repository-level coding tasks and specific harnesses. A voice model can be part of a coding-agent interface, but its relevant metrics would be transcription accuracy, latency, code/identifier handling and downstream task success when speech is used as input. The coding model's SWE-bench result remains a separate measurement.

Practical takeaway

The evidence supports a narrower, stronger conclusion than "Meta solved speech recognition."

Muse Voice Transcribe is a credible frontier streaming ASR release with an independently measured 3.1% final WER at roughly 0.16 seconds after detected end of speech on Artificial Analysis's current English-focused benchmark. Its unified architecture also performs endpointing and speaker diarization, while Meta reports 17.5% average DER on public multi-speaker datasets and supports 20+ speakers, hour-plus audio, 70+ trained languages and 25 extensively verified languages.

The limits are equally important. The AA-WER result is not a multilingual benchmark. The diarization result is vendor-reported over public datasets rather than independently reproduced in the same way as AA-WER. Early X praise is anecdotal, and public benchmark-selection criticism shows that specialist systems can be missing from launch charts. Pricing is attractive at a reported $3 per 1,000 minutes, but current billing and retention terms should be checked before production deployment.

For the next evidence frontier, the most valuable tests would be independent multilingual streaming WER by language, overlapping-speech diarization with strong specialist baselines, noisy-field audio, long-session drift, code/identifier transcription, and full voice-agent turn latency under fixed endpointing and network conditions.