Grok Voice Transcribe 2.0 vs Granite Speech 5.0 470M TurboCTC: 12,600× Real Time Meets Half a Second

Granite Speech 5.0 470M TurboCTC transcribes about 3.5 hours of audio every second on a single H200, and Grok Voice Transcribe 2.0 returns a first partial transcript 0.49 seconds after you stop talking. Those two numbers get placed side by side in comparison posts as if they were the same kind of measurement, and they are not remotely the same kind of measurement — one is a batched datacenter throughput figure with no latency attached, the other is a latency figure on a model whose throughput nobody has published. IBM released Granite Speech 5.0 470M TurboCTC on August 25, 2026 under Apache 2.0, dropping the language-model decoder entirely and decoding with a single non-autoregressive CTC pass. SpaceXAI shipped Grok Voice Transcribe 2.0 on September 18, 2026 as a managed API at $0.10 per audio hour for batch and $0.20 per hour for streaming. They are both excellent transcription models resolving opposite halves of the same problem, and picking between them is easier once you stop comparing the numbers directly.

12,600× real time, and the words that qualify it

The Granite figure is 12,600 RTFx measured on a single NVIDIA H200 with batched inference — IBM's number, reported at launch and not since independently reproduced. RTFx is a ratio of audio duration to processing time, so 12,600 means roughly 3.5 hours of audio per second of wall clock. IBM's own framing is that this is more than 20× the throughput of earlier Granite Speech releases, which is the claim that actually matters here: the speedup came from removing work, not from adding hardware.

What it is not is a latency number. A batched job that finishes 3.5 hours of audio per second can still take a long time to return the first token of any individual file, depending on how the batch is assembled and how much audio you feed it before it starts. IBM publishes no latency figures at all for TurboCTC. On the far-field leaderboard the two checkpoints are the two fastest entries, but throughput there was measured on a single NVIDIA L4, which is a datacenter-adjacent accelerator and not a phone.

The reason this matters is that the model is explicitly positioned for laptops, phones and edge devices — IBM's own framing — and nobody has published single-stream performance on any of those. Until someone does, treat "12,600×" as a statement about how cheaply you can chew through a backfill on rented GPUs, which is a genuinely valuable and well-evidenced claim, and not as a statement about how fast one user gets one sentence back.

What IBM removed, and what that costs you

TurboCTC is an encoder-only design: a 16-layer Conformer acoustic encoder trained with CTC, decoded greedily in one non-autoregressive pass, with a 16,384-unit subword output layer. There is no language model in the loop. That is where the speed comes from and it is also where the capability cuts come from, and they are not small ones. Compared with earlier Granite Speech models, TurboCTC drops speech translation, drops keyword biasing, and ships no timestamps or punctuation tooling.

Line that up against what the managed model includes at its list price and the shape of the trade becomes concrete:

• Key-term biasing — Granite Speech 5.0 470M TurboCTC has none vs up to 100 key terms per request on Grok Voice Transcribe 2.0

• Word-level timestamps — not shipped with TurboCTC vs included with confidence scores

• Speech translation — present in earlier Granite Speech models, removed here vs not offered either way

• Language coverage — English only vs automatic detection across a widely multilingual input set with formatting support in 25 languages

• Diarization — a separate problem you solve yourself vs included in the request price

• Channels — one stream per pass vs up to eight audio channels in one request

• Text normalization — none vs inverse text normalization turning spoken dates, currencies and phone numbers into written form

• Deployment — weights you download and run vs a managed endpoint in us-east-1

Read that list as a description of two different products rather than a scorecard. Removing diarization, biasing and normalization is what bought the throughput. A batch backfill of 40,000 call recordings into a search index does not need timestamps or speaker turns — it needs to be cheap and it needs to finish — and for that job the missing features are not missing, they are deleted weight.

The reverse is true for anything live. If you are building a voice agent, a key-term list is how you get "Kubernetes" and "Granola" and customer names spelled correctly, and a transcriber with no biasing mechanism will get them wrong every time with no way to fix it except fine-tuning, which is a different project with a different budget. That single missing feature disqualifies TurboCTC from some workloads and is irrelevant to others, which is why the model comparison is less useful than a workload comparison.

The accuracy comparison that cannot be made cleanly

IBM reports 5.00% aggregate word error rate for the Apache 2.0 checkpoint and 4.85% for the non-commercial sibling on the Open ASR Leaderboard, over public English short-form sets. Those are batch numbers on a board whose evaluations include AMI and Earnings22 and long-form conversational sets, and they are vendor-reported — run through the leaderboard's tooling but not yet absorbed into the official table, which also includes private test sets. The published per-set breakdown on the model card is worth reading, because the average hides a lot: LS Clean 1.42%, LS Other 2.66%, SPGISpeech 3.18%, Voxpopuli-Cleaned-AA 4.88%, Earnings22-Cleaned-AA-chunked 6.69%, AMI-Cleaned 7.52% and Gigaspeech-Cleaned 8.67%, averaging to exactly 5.00%. The same card quotes an average RTFx of 13,042.98 measured on one H200 — higher than the 12,600 figure IBM's launch post used, and the two numbers are both the company's, from the same week, on the same hardware.

Grok Voice Transcribe 2.0's 2.7% final-transcript figure is not a comparable measurement. It comes from Artificial Analysis's AA-WER Streaming board, a weighted composite of AA-AgentTalk at 50%, VoxPopuli at 25% and Earnings22 at 25% — roughly eight hours of agent-weighted English conversational audio, measured through a streaming interface. Different sets, different weighting, different mode of operation. Putting 2.7% next to 5.00% and concluding the managed model is roughly twice as accurate is the single most common mistake available in this comparison.

What can be said honestly is narrower and still useful. Both models are strong on the audio each was measured on. Grok Voice Transcribe 2.0 leads an independently-run streaming board that other models are also measured on, which makes its ranking checkable even if its exact value is not portable. Granite Speech 5.0 470M TurboCTC has an aggregate WER on the standard open ASR sets and, separately, a far-field result where the non-commercial checkpoint ranks fifth on accuracy and the Apache one ninth — measured on noisy and reverberant speech, which is the hardest condition in the set and arguably the most honest thing in either model's numbers.

If your audio is clean English read speech, the accuracy gap between these two is likely smaller than the leaderboard positions suggest. If it is noisy, accented, telephonic or non-English, neither board tells you much and your own test set is the only instrument that does.

Free weights against $1.67 an hour

The cost comparison is the one place these two models are genuinely easy to separate, and the number people usually cite is wrong in both directions.

• Batch transcription — Grok Voice Transcribe 2.0 at $0.10 per audio hour, or about $1.67 per 1,000 minutes vs Granite Speech 5.0 470M TurboCTC at no licence cost, plus GPU time

• Streaming — $0.20 per audio hour, about $3.33 per 1,000 minutes vs no hosted streaming path published by IBM

• Licence — a commercial API with no weights vs Apache 2.0 for the main checkpoint and CC-BY-NC-SA-4.0 for the more accurate non-commercial one

"Free" is doing a lot of work in that first row. A 470M encoder-only model is small enough to run on modest hardware — that is the point of the design, and the reason IBM trained it in ten days on eight H100s and shipped it against `transformers>=5.16.0` with a WebGPU demo — but somebody still pays for the GPU, the serving stack, the autoscaling, the retries and the on-call rotation. At small volumes the managed price is usually cheaper than the engineering time. At large steady volumes the rented-hardware path wins, and the crossover point is a function of your utilisation rather than your audio volume: a GPU you keep busy is cheap per hour, and one you keep warm for peak traffic is not.

One licensing detail is worth flagging before anyone architects around it. The more accurate of the two checkpoints, the one at 4.85% WER, is the non-commercial one under CC-BY-NC-SA-4.0. The Apache 2.0 checkpoint is the 5.00% one, and it is the one you may ship in a product. That is a 0.15-point accuracy difference traded for the right to use the model at all, and it also means any comparison that quotes 4.85% for "Granite Speech 5.0" and then discusses commercial deployment is quoting the wrong checkpoint.

The decision rule

Three questions separate these two models faster than any benchmark table.

Is the transcript consumed live, while the speaker is still talking? If yes, TurboCTC is not the candidate — not because it is slow, but because it has no streaming interface and no partial output, and a partial transcript is a different architectural commitment than a fast batch decode. If no, the throughput advantage becomes the whole story and IBM's model is very hard to beat on cost per hour of audio processed.

Does the work need domain vocabulary? If yes, a model with biasing wins by default, and the absence of key-term biasing in TurboCTC is disqualifying rather than merely inconvenient. If no, you are paying for a feature you will not use on the managed side.

Is the audio English read speech on decent hardware? If yes, both are strong and the choice is about operations. If no, both boards are uninformative and only your own evaluation matters.

Whichever way that lands, the operating layer is where this decision gets cheap. OrcaRouter puts 200+ models behind one OpenAI-compatible key with automatic failover across providers, so a single-region managed speech endpoint having a bad afternoon stops being your outage, and provider list price passes through at 0% markup — which means a vendor price change or a new checkpoint is live on our side the same day. For a workload where the right answer is genuinely unclear, that removes the cost of being wrong: benchmark both against your own audio behind one endpoint, keep the one that wins, and change the model string when the next one ships.

The short version: Granite Speech 5.0 470M TurboCTC is the better answer to "transcribe everything we have ever recorded, cheaply, on hardware we control," and Grok Voice Transcribe 2.0 is the better answer to "transcribe this as it happens, accurately, without us running the serving stack." The 12,600× headline and the 0.49-second headline are both true and neither one is evidence about the other.