VibeVoice-ASR-Streaming-7B vs Muse Voice Transcribe: Two Streaming Challengers, One Day Apart

Within roughly thirty-six hours in early September 2026, two big labs shipped streaming speech-to-text models that make the same headline promise — a live transcript that also tells you who said what, as the words are being spoken. On September 1, Meta Superintelligence Labs launched Muse Voice Transcribe as a hosted API priced at $3.00 per 1,000 audio-minutes, about $0.18 an hour, with streaming recognition, 20+ speaker diarization, and endpointing folded into one pass. On September 2, Microsoft Research uploaded VibeVoice-ASR-Streaming-7B to Hugging Face: MIT-licensed open weights, no announcement, a model card that claims streaming speaker-attributed transcription with hotwords and ten languages, and no hosted home anywhere. Same pitch, opposite business models, and evidence on both sides that is thinner than either lab would like you to notice.

This page is written as what-we-know-so-far, because neither model has an independently verified accuracy number. Muse Voice Transcribe's figures are Meta's launch-day claims, vendor-reported and unreproduced; VibeVoice-ASR-Streaming-7B has not published any streaming accuracy figure in text at all. The comparison below therefore leans on what is structural and knowable — architecture, price, license, documented behavior — and labels the few vendor numbers where they appear.

Two challengers to the batch pipeline, a day apart

Both models are attacks on the same assumption: that speech-to-text is something you run on a finished recording. Muse Voice Transcribe is the first product Meta calls a "real-time audio perception model" — one streaming pass that does recognition, speaker diarization across more than twenty voices, and endpoint detection, the job of deciding when a speaker has actually finished rather than paused. It ships on three surfaces at once: the Meta Model API at $0.18 an audio-hour, Meta AI for Mac where holding the Fn key dictates into any application, and Muse Code. Microsoft's VibeVoice-ASR-Streaming-7B is the streaming member of the open VibeVoice family, and its release trail is much quieter: a Hugging Face repository created September 2 (timestamps show 15:46 UTC, last touched three minutes later), eight safetensors shards under MIT, and a news-log line dated September 3 in the microsoft/VibeVoice GitHub repository that calls it "a unified streaming ASR model that continuously transcribes who said what as speech arrives, with support for customized hotwords and 10 languages." A day after upload it still showed zero downloads.

The feature collision

• Streaming mechanism — Muse Voice Transcribe consumes audio in 80-millisecond chunks and commits words as they stabilize vs VibeVoice-ASR-Streaming-7B processes ~2.9-second chunks with ~0.5 seconds of lookahead, emitting text once per resolved chunk.

• Speaker attribution — 20+ speakers in one pass, vendor-reported 17.5% average diarization error vs streaming who-said-what output claimed on the model card, unverified.

• Endpointing — built-in: the stream carries a finished-utterance boundary vs not documented as an output signal.

• Context controls — language, keyword, and context biasing vs customized hotwords via a context prompt (--context_info).

• Languages — trained on 70+, 25 extensively verified at launch, native code-switching vs a declared 10 (en, zh, es, pt, de, ja, ko, fr, ru, it).

• Evidence — launch-day vendor claims: 3.1% WER on finals at 0.16 s after speech, 3.6% on first partials, citing the Artificial Analysis streaming leaderboard vs no streaming WER or latency figure published as text.

• Cost and license — $0.18 per audio-hour, hosted, closed weights vs $0 for the weights (MIT), roughly 18 GB of bf16, self-hosted.

Two very different ideas of "streaming"

The word streaming covers a wide cadence range, and the gap between these two is wide enough to change what you can build on each. Muse Voice Transcribe streams at word granularity. Meta's architecture consumes audio in 80-millisecond chunks and uses what it calls "adaptive delay" — a delay policy learned through reinforcement learning that commits each word as soon as the model is confident, holding more context for words it is less sure about rather than applying one fixed latency budget. The vendor claims final transcripts 0.16 seconds after the speaker stops and first partials 0.13 seconds in. That cadence is built for interactive loops: live captioning, dictation, a voice agent that needs to respond to a finished utterance almost immediately.

VibeVoice-ASR-Streaming-7B streams at a coarser grain. Its preprocessor config shows audio arriving at 24 kHz, compressed 3,200× into tokens at about 7.5 frames per second, then processed in chunks of 22 frames with a 4-frame lookahead — roughly 2.9 seconds of audio per chunk, about half a second of lookahead, figures derived from the config rather than measured latency. Text is emitted as each chunk resolves, with context from earlier chunks preserved through the KV cache so a long session does not recompute from scratch. A transcript that grows in roughly three-second increments is fine for meeting notes and call analytics; it is a different product from sub-second word streaming for live conversation. Neither lab has published an end-to-end latency measurement for the streaming checkpoint, so treat the cadence as the design and the real-world feel as untested.

Speaker labels and endpointing: built in on one, claimed on the other

Speaker attribution is where streaming products most often overstate, and both of these make the claim. Muse Voice Transcribe's version is concrete and structural: diarization runs inside the same streaming pass as recognition, so a recording of a twenty-person call can carry per-speaker labels without a separate "upload, wait, get speakers back" step, and Meta reports an average diarization error rate of 17.5% — a vendor number, unreproduced, but a number attached to a real mechanism. It also emits endpoint boundaries, the quieter capability: an application gets a clean "this speaker's turn is done" signal without building a voice-activity detector, which is why voice agents built on raw ASR so often cut people off.

VibeVoice-ASR-Streaming-7B claims streaming who-said-what output on its model card, in line with the batch VibeVoice-ASR's structured Who/When/What output — but for the streaming checkpoint the claim is unverified, no independent test has reproduced it, and there is no documented endpointing signal of the kind Muse ships. If your workload is genuinely multi-speaker live audio, Muse is offering a more complete mechanism today; if you are evaluating whether an open-weights model can do the same job on hardware you control, the VibeVoice claim is exactly the kind of thing to test with your own meeting recordings before believing it.

The evidence: launch-day claims against a blank card

It is worth being precise about what each side has, because neither has what a buyer actually wants. Muse Voice Transcribe has a dense set of vendor numbers — 3.1% word error rate on final transcripts, 3.6% on first partials, 0.16-second final latency, 17.5% average diarization error — all published by Meta on launch day and pointing at the Artificial Analysis streaming speech-to-text leaderboard, where Meta claims the top spot. Those are claims, unreproduced by anyone outside the lab, but they are specific enough to falsify, which is a kind of progress.

VibeVoice-ASR-Streaming-7B has none of that. Its model card ships an evaluation figure as an image and the streaming technical report is a PDF, but there is no quotable streaming WER or latency figure in prose. The only numeric anchor in the VibeVoice family is the batch VibeVoice-ASR card's vendor-reported table — 7.77% average WER across eight English test sets, 2.20% on LibriSpeech clean — which is explicitly not this checkpoint's number. When a launch-day claim meets a blank card, the honest tie-breaker is everything except the headline WER: price, license, streaming cadence, and the features each side actually documents.

Languages: 25 verified against 10 declared

Language coverage separates the two more than the headline accuracy claims do. Muse Voice Transcribe was trained on 70+ languages but Meta validates 25 at launch — and the verified list, not the training list, is the production coverage, with native code-switching as the differentiator: it is designed to switch languages mid-sentence without being told. VibeVoice-ASR-Streaming-7B declares 10 languages in its model card — English, Chinese, Spanish, Portuguese, German, Japanese, Korean, French, Russian, Italian — against the 50+ of the batch VibeVoice-ASR. If your traffic includes code-switched conversation, Muse has the more specific pitch; if it sits inside those ten languages, the Microsoft model's coverage is at least explicit, which is more than most launch materials offer.

Cost and what you keep

The business-model difference is the cleanest way to decide. Muse Voice Transcribe at $0.18 per audio-hour undercuts every major streaming transcription API on list price — no GPU, no ops, closed weights, and the price is set by Meta. VibeVoice-ASR-Streaming-7B costs nothing to download but roughly 18 GB of bf16 weights before KV cache to self-host on a 24 GB-class GPU, with the microsoft/VibeVoice demo scripts or the vLLM plugin as the documented serving paths, and no inference provider hosting it yet. The MIT license is the durable asset: the weights are yours to keep and fine-tune in a way no API subscription is. That is also where the pricing picture could get interesting — the moment a provider does host the open checkpoint, the $0.18-an-hour question becomes a market price rather than a single vendor's list price, which is exactly the situation where a pass-through router earns its keep. OrcaRouter does not route speech-to-text today, and neither of these models is in its catalog; what it does do is pass provider list prices through at 0% markup on the 200+ language models that consume a live transcript, so a vendor price cut anywhere in that stack is live on the buyer's side the same day it is announced.

Frequently asked questions

Does Muse Voice Transcribe's 3.1% WER mean it is more accurate than VibeVoice-ASR-Streaming-7B?

No, and the comparison is not meaningful yet. Meta's 3.1% is a launch-day claim for the streaming path, vendor-reported and unreproduced. VibeVoice-ASR-Streaming-7B has published no streaming accuracy figure in text at all. Until both models have been run through the same public harness on the same audio, the only honest statement is that Muse has a specific claim that can be tested and VibeVoice does not yet have one.

Can I use either model on audio that is already recorded?

Yes to both, with different shapes. Muse Voice Transcribe handles recordings longer than an hour through the same streaming API. VibeVoice-ASR-Streaming-7B's vLLM plugin exposes a whole-file transcribe endpoint alongside its WebSocket stream endpoint. But both are designed and priced for the live case — Muse by its streaming-only business model, VibeVoice by the chunked architecture that emits as audio arrives — so if your workload is purely batch files, a dedicated file transcription API will usually be cheaper and better measured than either.