Nemotron 3.5 Transcribe vs Muse Voice Transcribe: One Processes the Meeting, the Other Sits In It

Two recordings decide this comparison before any benchmark does. One is a call happening right now, where the transcript has to keep up with the conversation and a two-second lag makes the tool useless. The other is a two-hour archive of last quarter's customer calls, where nobody cares whether the words arrive in 200 milliseconds or twenty minutes. NVIDIA Nemotron 3.5 Transcribe, released on October 28, 2026 as an open-weight offline checkpoint, is built for the second recording. Meta Muse Voice Transcribe, shipped by Meta Superintelligence Labs on September 1, 2026, is built for the first. Most comparison pages put their word error rates in one column and call it a day, which is exactly the comparison that cannot be made.

The clock, not the word error rate

Meta's model is a streaming perception layer. It consumes audio in 80-millisecond chunks, emits soft tokens, and uses a reinforcement-trained "adaptive delay" mechanism to decide word by word how much context to wait for before committing to text. Meta's launch material, as reported, puts the gap from end of speech to final text at roughly 0.16 seconds, and the model is served over a WebSocket for live audio or an HTTP endpoint for one-shot files. It also handles endpointing and voice-activity detection itself, so the pipeline around it is thin.

NVIDIA's model is the opposite shape. The card says it plainly: "This is an offline model," and adds that internal chunking does not make it a streaming service. You hand it a file of up to one hour and it returns a transcript. There is no partial output, no interim hypothesis, no barge-in. NVIDIA sells a separate 0.6B streaming recogniser for live audio; that is a different checkpoint, not a mode of this one.

So the headline numbers are measured on different instruments. Meta reports 3.1% word error rate and first place on Artificial Analysis's streaming speech-to-text leaderboard. NVIDIA's card reports a 4.71% mean across the nine public English Open ASR sets, while Artificial Analysis's non-streaming table shows 3.4% for the deployed endpoint — and its FAQ says 2.8% for the same model. Those are three different measurements of two different tasks. The 3.1% is a streaming result; the 3.4% is a batch result on a different audio distribution. Subtracting them produces a 0.3-point "win" that means nothing, and the streaming model's disadvantage on long, noisy, accented audio is real but invisible in that arithmetic.

Speakers: twenty-something against eight

The diarization gap is the one place where the numbers are measuring roughly the same thing, and it is large.

• Speaker ceiling — Meta advertises 20+ speakers in a single recording; NVIDIA trains on mixtures up to eight, and its model card states no inference-time maximum.

• Overlap — Meta's own page qualifies that diarization applies to "non-overlapping speech." NVIDIA's whole architectural argument is overlapping speech: the frozen diarizer's speaker activity is injected into the encoder so the model can attribute words while people talk over each other. That is the claim behind its 15.80 average cpWER on NVIDIA's multi-speaker set.

• Language coverage for labels — Meta serves 25+ languages; NVIDIA supports 40 for transcription but marks 13 of them for multispeaker output. A Turkish meeting transcribes and does not get labels.

• Label format — Meta emits bare speaker letters such as "A"; NVIDIA emits <spk:0>-style tags that map back to diarization segments.

If your use case is a panel, a courtroom or a support call with three overlapping participants, the eight-speaker ceiling is the constraint that decides it, not the WER column. If your use case is a twenty-person all-hands where people mostly take turns, Meta's 20+ is the number that matters and NVIDIA's eight is a hard wall.

Languages: 25 against 40, and neither number is what it looks like

Meta's own developer page says "25+ languages" and supports code-switching within and between sentences, which matters for bilingual meetings. Launch coverage reported that the model was trained on 70+ languages with 25 extensively validated — a useful distinction, because a language that is "trained" but not validated is a language you should test before trusting. NVIDIA's card lists 40 languages with ASR support, 31 with bidirectional translation to and from English, and a separate 48-label language-ID component.

• If you need European coverage breadth — Meta's 25 validated languages are narrower than NVIDIA's 40, but Meta's list leans toward the languages enterprise customers actually record.

• If you need translation — only NVIDIA does it. Meta's model transcribes; it does not translate the transcript into another language.

• If you need code-switching — Meta claims it natively; NVIDIA's card notes nothing comparable, and its audio understanding is English-only regardless of the transcription language.

What the transcript actually contains

Timestamps are where a cheap-looking difference becomes an expensive one. Meta provides turn-level timestamps — segment boundaries, not per-word times — and does not expose confidence scores at all. NVIDIA provides word timestamps through a CTC adapter, with the caveat that timestamps are unavailable when the DFlash2 speculative-decoding draft is enabled.

• Timestamps — Meta turn-level; NVIDIA word-level, but not alongside its fastest decode path.

• Confidence — Meta publishes none; NVIDIA does not publish per-word confidence either.

• Input format — Meta requires mono 16-bit PCM at 16 or 24 kHz and tells you to convert everything else with ffmpeg; NVIDIA accepts 16 kHz mono and handles the file directly.

• Prompting — both bias the output. Meta's keyword and language biasing "only nudge the model and do not guarantee spelling," per its docs. NVIDIA's contextual prompting reports named-entity WER falling from 12.21% to 2.20% with a fine-grained entity list — vendor-run, but a much stronger claim, and one that implies a different mechanism than a soft nudge.

• Output styles — NVIDIA's five styles (verbatim, written-form, contextual, speaker-attributed, rich) are selected by instruction. Meta returns one format and expects you to post-process.

Price per audio hour

Meta charges $3.00 per 1,000 minutes, which is $0.18 per hour of audio, with no tiers or free allowance published. NVIDIA's hosted listing on Artificial Analysis is $2.00 per 1,000 minutes — about $0.12 per hour — and the weights are free to download under OpenMDW-1.1, so self-hosting is a third option that Meta does not offer at any price.

The arithmetic at volume is straightforward and does not favour the cheaper sticker. A thousand hours of audio per month costs $180 through Meta and $120 through a hosted Nemotron endpoint. Self-hosted, the same thousand hours is under an hour of GPU time on an RTX PRO 6000 Blackwell at NVIDIA's NVFP4 throughput claim — call it $2 of compute at a rented-GPU rate, plus operations, storage and the engineer who owns it. At 10,000 hours a month the gap between $1,800 and a few hundred dollars pays for a lot of operations; at 100 hours a month it pays for none of it.

What self-hosting actually buys is not the price. It is the ability to run transcription on audio that is contractually forbidden from leaving your network, and the ability to keep the pipeline alive when a vendor's endpoint degrades. Those are the reasons regulated teams pay for GPUs they could otherwise rent.

Where each one breaks

Meta's failure modes are latency-adjacent. Its diarization assumes non-overlapping speech, so a heated support call with interruptions will mislabel turns. Its transcript has no word timestamps, which breaks subtitle generation, audio-search and any downstream alignment to video. Its input contract is strict, and its biasing is advisory. It has no translation, no audio question answering, and no self-hosted option.

NVIDIA's failure modes are workflow-adjacent. It cannot stream, so it cannot serve live captions, voice agents or real-time call routing. Its audio understanding is English-only. It caps at one hour per recording. Its shipped SGLang-Omni integration omits speaker-attributed output, word timestamps and translation, so some of what the card promises requires the vLLM path or the NIM container. And it can occasionally fail to terminate, which the card recommends guarding against.

Neither set of failures is a bug in the other model's favour. They are the two shapes of the problem: a model that is present while the audio happens, and a model that is thorough after it ends.

If you need both, the pipeline is not a compromise

The most common real deployment is not a choice at all. Live audio goes through Meta's streaming endpoint, because that is the only way to get text while the call is happening — captions, agent assist, live search. The same recording is then re-transcribed offline by NVIDIA's checkpoint overnight, because the batch pass produces cleaner text, speaker-attributed word timestamps and, where needed, an English translation of a non-English call. The streaming transcript is for the moment; the batch transcript is for the record. Teams that try to force one model into both roles usually end up with either a lagging live view or an expensive streaming bill for audio nobody is watching.

The downstream half of that pipeline — summarising calls, scoring agent performance, redacting personal data, routing flagged conversations to a human — is ordinary text work, and it does not belong on the same GPU as the ASR. OrcaRouter runs those text calls behind one key with 0% markup, passing the provider's list price through unchanged, with automatic failover when a provider wobbles and a routing DSL that can send summarisation and classification to different models inside a single workflow. Neither Meta Muse Voice Transcribe nor NVIDIA Nemotron 3.5 Transcribe is in OrcaRouter's catalogue; the router sits after the transcript, where the text workload lives.

Three questions settle this faster than any leaderboard. How long after the audio arrives does the text need to exist? If the answer is "during the call," the decision is Meta's and the WER discussion is decoration. How many people talk at once, and do they overlap? If the answer is "more than eight, and yes," that is NVIDIA's terrain. And is the audio allowed to leave your network? If the answer is no, only one of these two models can be deployed at all.

Everything else — the 3.1%, the 4.71%, the $0.18 — is a tiebreaker between two products that were designed for different moments in the same meeting.