Muse Voice Transcribe vs Granite Speech 5.0 470M TurboCTC: GPU You Own or API You Meter

The most useful thing to know about this matchup is that Muse Voice Transcribe and Granite Speech 5.0 470M TurboCTC are barely substitutes. IBM shipped Granite Speech 5.0 470M TurboCTC on August 25, 2026 as a 470-million-parameter, Apache-2.0, English-only ASR model built for self-hosting — an encoder-only Conformer trained with CTC that IBM says transcribes at over 12,600× real-time on a single NVIDIA H200. Meta Superintelligence Labs shipped Muse Voice Transcribe on September 1, 2026 as a proprietary, hosted, streaming audio-perception model — 25 languages verified at launch, 20+ speaker diarization, endpointing, $3.00 per 1,000 audio-minutes. One wants your GPU; the other wants your API key. Comparing their headline WER numbers would miss the point entirely — the real question is where the transcription is allowed to happen, and what each model can do once it gets there.

Two deployment philosophies

Granite Speech 5.0 470M TurboCTC is the apotheosis of the open-weights edge ASR movement. At 470M parameters it fits comfortably on a laptop, a phone, or a single GPU; the license is Apache-2.0 so you can embed it in a commercial product; and IBM ships a WebGPU demo that runs transcription live in a browser tab with no server at all. The architecture is deliberately stripped down — 16 Conformer blocks, greedy decoding, no language-model backbone — because the entire design goal is throughput on modest hardware. Muse Voice Transcribe is the opposite bet: a large proprietary model you never see, served from Meta's infrastructure at $0.18 per audio-hour, with the value delivered in features rather than in silicon. If your constraint is "no audio leaves this building," the choice is already made and it is Granite. If your constraint is "I want the most capable streaming model and I will pay per minute," it is Muse.

Speed means different things here

Granite's headline, 12,600 RTFx, is a throughput figure: one H200 running batched inference plows through 12,600 seconds of audio per second — three and a half hours of audio every second. That is the metric that matters for a call-center backlog, a media archive, or any batch pipeline where you feed the GPU continuously. It is not a latency figure, and single-stream interactive latency is not where this model shines. Muse Voice Transcribe is the mirror image: its 80-millisecond chunk processing and adaptive delay are built for the first 300 milliseconds after a speaker pauses, not for sustained bulk throughput. On the metric each was designed for, both are strong; on the other's home turf, both are out of their element. If you need to transcribe a million hours of English archive audio, Granite's raw economics win by an order of magnitude. If you need a live multi-speaker conversation turned into labeled text as it happens, Muse is the one that even offers that.

Accuracy, labeled honestly

• Granite Speech 5.0 470M TurboCTC — 5.00% aggregate WER on the Open ASR leaderboard's public English short-form test sets; IBM's own run of an official harness, vendor-reported and unreproduced, on English audio only.

• Muse Voice Transcribe — 3.1% streaming WER, Meta's launch-day claim citing the Artificial Analysis streaming leaderboard; 3.6% on first partial transcripts; also vendor-reported and unreproduced.

The two figures are not directly comparable — different corpora, one English-only and offline, one multilingual and streaming — which is exactly why this matchup should not be decided on WER alone. On the evidence available, both are plausible but unverified for their respective workloads. What is structurally certain is that Granite's number covers English only, while Muse's claim sits in a streaming, code-switching context where Granite cannot even compete.

What each model simply cannot do

Granite Speech 5.0 470M TurboCTC drops everything that makes a transcript useful to a product team. It is English-only. It has no speaker diarization — every voice lands in one stream. It does not emit punctuation, capitalization, or timestamps, and the CTC design gives up keyword biasing and speech translation that earlier Granite Speech models had. Those are not quality gaps; they are architectural choices, and they mean a production stack built on Granite still needs a punctuation model, a diarizer, and a VAD layer bolted on. Muse Voice Transcribe is where those features live in the model itself: 20+ speaker diarization and endpointing stream out alongside the words, and the language coverage with mid-sentence code-switching is something Granite has no answer for at all. The honest summary: Granite is a recognition engine that assumes you will build the product; Muse is a product-shaped API that assumes you want a transcript, speakers, and turn boundaries on arrival.

License and ownership

Granite Speech 5.0 470M TurboCTC is Apache-2.0, and a non-commercial sibling (granite-speech-5.0-470m-turboctc-nc, CC-BY-NC-SA-4.0) exists for research use with a slightly better 4.85% WER thanks to extra training data. Apache-2.0 means you can serve it, embed it, fine-tune it, and sell a product around it with no royalties and no audit risk. Muse Voice Transcribe is closed and hosted-only: no weights, no self-host, no fine-tuning, and your transcript data flows through Meta's service under Meta's terms. For regulated industries, or any team whose data policy forbids third-party audio processing, that single fact ends the comparison before the benchmarks start. For everyone else, "Apache-2.0 and your own GPU" and "proprietary and metered" are simply different risk profiles, not better and worse.

The cost question

Granite at 12,600 RTFx makes the self-hosted math almost silly for batch English audio: one H200-hour of compute covers over 12,000 hours of audio, so a thousand hours of transcription costs a few dollars of raw GPU time at typical rental rates — before you account for idle time, engineering, and the missing diarization layer. Muse Voice Transcribe bills $0.18 per audio-hour, which is cheap for a streaming API but not cheap next to a fed GPU. The honest rule of thumb: if the audio is English, pre-recorded, and you have volume, self-host Granite and the economics win. If the audio is live, multi-speaker, or multilingual, Muse is priced as a managed feature service — and $180 for a thousand hours of live, diarized, endpointed streaming is a small number relative to what that stack costs to build yourself.

Running both without running a mess

Teams that need both profiles — self-host Granite for the English batch lane and a managed API for live multilingual conversations — end up with two very different integrations. The hosted half is precisely the shape of workload a routing gateway serves: one API key across many providers, list prices passed through at 0% markup so the vendor's per-minute figure is the number you actually get billed, and automatic failover if a provider endpoint degrades. The self-hosted half stays yours. There is no need to route Granite through a gateway — it lives on your own hardware — but the mixed stack it implies is exactly the kind of thing a single-API platform exists to keep from becoming two parallel engineering projects.

Which one should you pick

If your job is English transcription at scale — archives, call recordings, subtitle pipelines — and you have a GPU to point at it, Granite Speech 5.0 470M TurboCTC is the better engineering decision on cost, license, and control, with the caveat that you will build the punctuation, diarization, and streaming layers yourself. If your job is live, multi-speaker, or multilingual conversation — voice agents, live captions, meeting products — Muse Voice Transcribe is offering features Granite does not have, at a managed-API price, with the caveat that its numbers are launch-day claims on closed weights. They are not competitors so much as answers to different questions, and the teams that get this right will often end up running both.