Grok Voice Transcribe 2.0 vs VibeVoice ASR Streaming 7B: The Number Microsoft Didn't Publish

VibeVoice ASR Streaming 7B is Microsoft's unified streaming speech recogniser, uploaded to Hugging Face on September 2, 2026 and announced in the microsoft/VibeVoice repository a day later under the MIT licence, and it does something neither Grok Voice Transcribe 2.0 nor most of its competitors do: it emits speaker-attributed text as audio arrives, with no separate diarization stage. Microsoft's technical report says the 7B checkpoint achieves the lowest average WER and CER across five evaluation sets and the best or tied-best speaker attribution in 12 of 13 evaluation settings. What the model card does not do is publish a single number in text — no WER, no RTF, no latency figure; the results exist only as images in the report. Grok Voice Transcribe 2.0, shipped sixteen days later on September 18, 2026 at $0.10 per audio hour for batch and $0.20 per hour for streaming, publishes everything: 2.7% final-transcript WER and 3.4% first-partial WER on Artificial Analysis's streaming board, 0.49 seconds to each. So the honest starting point for this comparison is that one of the two models is measurably better documented than the other, and that asymmetry is itself the most decision-relevant fact in the matchup.

What Microsoft published, and what a reader can do with it

The VibeVoice report is real research and the claims in it are specific in shape, if not in value. It evaluated four meeting benchmarks — AliMeeting, AISHELL-4, AMI-SDM and AMI-IHM — plus MLC-Challenge, across nine languages, and reports two headline results: the 7B model had the lowest average WER/CER across the five sets, and the best or tied-best speaker attribution in 12 of 13 evaluation settings. Both are relative claims, which is a meaningful kind of claim: "lowest across these five sets" is a statement about ordering, and ordering is what a buyer needs.

What is absent is every absolute figure. There is no WER percentage to compare against anything, no throughput number, no latency measurement, and no per-set breakdown in text. A comparison table that puts VibeVoice next to Grok Voice Transcribe 2.0 and assigns the Microsoft model a number is inventing that number. What can be said is that VibeVoice won its own five-set evaluation and that nobody outside Microsoft has re-run it — no independent benchmark, no third-party evaluation, and at the time of writing no hosted inference provider listing the model at all. The comparison is therefore between a documented number and an undocumented ranking, and the undocumented ranking is not the weaker of the two just because it lacks a decimal point.

Grok Voice Transcribe 2.0's documentation has the opposite problem. Its 2.7% and 3.4% come from Artificial Analysis's AA-WER Streaming board, a weighted composite of AA-AgentTalk at 50% with VoxPopuli and Earnings22 at 25% each — about eight hours of agent-shaped English conversational audio, independently run, which is a genuinely stronger provenance than a vendor's own evaluation. But it is one narrow slice of English, and the board page itself plots 27 of the 33 models SpaceXAI counts when it claims first place of 32. A precise number on a narrow test and an imprecise claim on a broader one are not directly comparable, and this article will not pretend otherwise.

Two and a half seconds against half a second

The latency comparison is the one place where the two models can be put side by side without any caveat about datasets, because both publish their streaming configuration — and the difference is architectural rather than a tuning choice.

VibeVoice ASR Streaming 7B works in chunks. Its released checkpoints use 22 latent frames per chunk, which at the model's 7.5 latent frames per second is about 2.9 seconds of audio per chunk, with a lookahead of four latent frames — roughly 0.5 seconds of future audio — and an expected speaker-attribution latency of about 2.00 seconds. It emits text once per chunk. That is a real streaming system, but the cadence is measured in seconds, not milliseconds.

Grok Voice Transcribe 2.0 returns its first partial transcript 0.49 seconds after the speaker stops, and finalises 0.49 seconds after end of speech. It bought that speed with accuracy — the first-partial error rate fell from 18.3% in version 1.0 to 3.4% in 2.0, at the cost of going from 0.25 seconds to 0.49 seconds on the same measure.

Put next to each other:

• Streaming cadence — Grok Voice Transcribe 2.0 emits partials continuously with a 0.49-second first-partial latency vs VibeVoice ASR Streaming 7B emitting one chunk of text roughly every 2.9 seconds

• Speaker attribution latency — included within the same 0.49-second path vs approximately 2.00 seconds by design

• Lookahead — not published vs four latent frames, about 0.5 seconds of future audio

• Practical effect — a voice agent can act on a sentence in progress vs a system that receives a speaker-labelled paragraph every three seconds

Neither of these is wrong. A 2.9-second chunk is a perfectly reasonable design for meeting transcription, where the output is a document and the value is that the document arrives during the meeting rather than after it. It is the wrong design for a conversational agent, where a three-second silence is not a pause, it is a failure. The gap between the two cadences is roughly six-fold, and it maps almost exactly onto the difference between transcribing a conversation and participating in one.

Speaker attribution as an output, not a pipeline stage

This is where the Microsoft model is genuinely ahead, and the design is worth understanding because it is the reason the chunking is coarse.

VibeVoice uses two pre-trained tokenizers — an acoustic encoder and a semantic encoder — running at 24 kHz with 3,200× downsampling to produce 7.5 latent frames per second, one every 133 milliseconds. Those frames are projected into a Qwen2.5 language-model backbone, and incoming speech and generated text are interleaved as a single sequence, with previously observed speech and text retained in the model's context. The consequence is that speaker identity is never a separate problem: the model is not transcribing and then deciding who spoke, it is generating text conditioned on an audio history that still contains the voices. That is why there is no diarization stage to align, and why Microsoft's claim about speaker attribution across 12 of 13 settings is architecturally credible rather than a bolt-on result.

The cost of that design is history retention that grows linearly with recording length, which is why the released checkpoints target recordings of up to eight minutes. That is a real ceiling and it is stated plainly. It rules the model out for long-form work — a two-hour earnings call, a full day of contact-centre audio — unless you build your own segmentation and re-initialisation, which reintroduces exactly the complexity the architecture was designed to remove.

Grok Voice Transcribe 2.0 solves the same problem differently and with a different set of limits. It includes diarization at no extra cost and supports up to eight independent audio channels in a single request. For contact-centre telephony, where each participant often arrives on their own channel, channel separation is more reliable than diarization and has no error rate attached. For mixed audio it depends on the quality of the diarization, and unlike Meta's Muse Voice Transcribe — which published a 17.5% average diarization error rate — SpaceXAI has not published a diarization figure at all. So both models leave a gap in the diarization evidence: one publishes a ranking without a value, the other publishes a feature list without a measurement.

Eighteen gigabytes, ten languages, MIT

The deployment facts diverge so completely that they are almost not a comparison. VibeVoice ASR Streaming 7B is roughly 18 GB of bf16 weights — the Hugging Face card lists 9B total parameters for the 7B-named checkpoint, with a 1.5B sibling at about 5.6 GB — MIT licensed, with Python demos and a vLLM plugin for serving. Ten languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. Microsoft positions it for research and development.

Grok Voice Transcribe 2.0 is a managed endpoint with no weights to download, 12 input formats from 8 kHz telephone audio to 48 kHz, up to eight channels, 100 key terms per request, automatic language detection with mid-recording switching, inverse text normalization across 25 languages, files up to 500 MB, and service limits of 10 requests per second and 100 concurrent streaming sessions per team. It runs in us-east-1 and only us-east-1.

• Weights — MIT, 18 GB, yours to run anywhere vs none, vendor-hosted only

• Languages — 10 named languages vs automatic detection with formatting support across 25

• Key-term biasing — supported via hotwords vs up to 100 key terms per request

• Recording length — about eight minutes per checkpoint before history retention becomes the constraint vs no published ceiling, files to 500 MB

• Region control — whatever you deploy on vs us-east-1

• Cost — no licence fee, plus 18 GB of GPU memory and a serving stack vs $0.10 per batch hour and $0.20 per streaming hour

An 18 GB model is not something you run on a laptop, and the vLLM plugin means a GPU with real memory. Against that, an MIT licence on a model of that size is unusual and valuable: it is the only one of the two you can run inside your own network with no vendor relationship at all, which for regulated audio is not a preference but a requirement. The managed alternative is cheap per hour and impossible to deploy on-premises.

Where the routing decision belongs

Cost is where the two models finally become comparable in a way that produces a number. Grok Voice Transcribe 2.0's batch path costs $0.10 per audio hour, about $1.67 per 1,000 minutes, and its streaming path $0.20, about $3.33. VibeVoice ASR Streaming 7B has no licence cost at all; what it has instead is roughly 18 GB of resident GPU memory and a vLLM serving stack you operate. A 7B-class model at that size is not going to be cheap per hour of audio on rented hardware, and the utilisation curve is unforgiving — a GPU held warm for peak traffic costs the same whether it is transcribing or idle.

That is the usual shape of the build-versus-buy argument, and it resolves differently depending on volume and on whether the audio is allowed to leave your network. What has changed in the last month is how fast the answer moves: the streaming accuracy leader changed hands twice in three weeks, and a model released in early September was displaced by mid-September. Any architecture that makes the model choice expensive to reverse is now the wrong architecture for this category.

OrcaRouter is where that reversal gets cheap. One OpenAI-compatible key covers 200+ models with automatic failover across providers, so a single-region managed endpoint failing is a routing event rather than an outage, and provider list price passes through at 0% markup — meaning a vendor price change or a new checkpoint is live on our side the same day. For a team that wants to test VibeVoice against Grok Voice Transcribe 2.0 on its own audio, or to keep a self-hosted fallback behind the same interface as a managed primary, the call site stops being the thing that has to change.

The honest verdict: VibeVoice ASR Streaming 7B is the more interesting model and Grok Voice Transcribe 2.0 is the more usable product, and the gap between those two statements is entirely explained by one vendor publishing charts and the other publishing numbers. If your requirement is on-premises speaker-attributed transcription of meetings under eight minutes long, the Microsoft checkpoint is the only one of the two that qualifies. If your requirement is a voice agent that answers within a conversational pause, a 2.9-second chunk cadence rules it out before accuracy is discussed. And if you are waiting for the comparison that would settle it — the same audio through both models, scored by someone who does not work for either company — that measurement does not exist yet, which is worth remembering the next time a table puts a number next to VibeVoice's name.