Grok Voice Transcribe 2.0 vs NVIDIA Parakeet TDT 0.6B: Two Leaderboards, One Misleading Comparison

NVIDIA Parakeet TDT 0.6B is a 600-million-parameter FastConformer-TDT transducer released on August 14, 2025 under CC-BY-4.0, and it has been the default recommendation for anyone who wants to run transcription themselves for over a year. Grok Voice Transcribe 2.0 is SpaceXAI's managed speech-to-text model, shipped September 18, 2026, priced at $0.10 per audio hour for batch and $0.20 per hour for streaming, with a 2.7% final word error rate on Artificial Analysis's streaming board. If you search for a comparison between them you will find Parakeet quoted at 13.7% WER and Grok Voice Transcribe 2.0 at 2.7%, which makes the managed model look five times more accurate and is a genuinely misleading thing to read. Those two numbers come from different Artificial Analysis indices, built on different datasets, published years apart, and one of them has been superseded by the vendor's own newer measurements. The real comparison is more interesting, and it is about what 600 megabytes of weights buys you that an API cannot.

The 13.7% number, and why you should not use it

Artificial Analysis's Q3 2025 State of AI report placed NVIDIA Parakeet TDT third on its AA-WER index at 13.7%, behind Google's Chirp 2 at 11.6% and NVIDIA's own Canary Qwen at 13.2%. That index was an audio-duration-weighted average across roughly two hours each of VoxPopuli, Earnings-22 and AMI-SDM — real-world speech with varied accents, domain vocabulary and difficult channel conditions, which is why every number on it is high. A 13.7% WER on AMI-SDM, which is far-field meeting audio recorded in reverberant rooms, is not a bad result; it is a hard test.

That index no longer exists in that form. Artificial Analysis rebuilt it as AA-WER v2, weighted AA-AgentTalk at 50% with VoxPopuli-Cleaned-AA and Earnings22-Cleaned-AA at 25% each, about eight hours of audio, and the cleaning and reweighting moved every model's number substantially. On the current non-streaming board, Parakeet TDT 0.6B V3 sits in the low single digits — the same band as other open-weights leaders — rather than anywhere near 13.7%, and the top of the board is down around 2%. Anyone still quoting 13.7% for Parakeet is quoting a retired measurement, and if they are comparing it against a 2026 streaming figure they are also comparing across two different boards.

What can honestly be put side by side is narrower. NVIDIA's own evaluation on the Open ASR Leaderboard — the standard public English short-form sets, run through that board's tooling — reports a mean WER of 6.32% for Parakeet TDT 0.6B V3 with an RTFx of 3,332.74, and a mean of 6.05% for the English-only v2. Grok Voice Transcribe 2.0's 2.7% is a final-transcript figure on the streaming board, measured after end of speech, on agent-weighted English conversational audio. Different sets, different board, different mode. The managed model is very likely more accurate on the audio the two boards share; the size of that advantage is not 5×, and nobody has published the measurement that would settle it.

What the two models are actually shaped like

The architectural difference explains most of the practical difference. Parakeet TDT 0.6B is a FastConformer encoder with a token-and-duration transducer decoder — it predicts both a token and how many frames to advance, which lets it skip rather than emit blanks, and that is the main source of its efficiency. It ships with automatic language detection across 25 European languages, word- and segment-level timestamps, punctuation and capitalization, and it handles long audio — up to 24 minutes with full attention, or around three hours with local attention. It is served through NeMo 2.2, has an MLX path for Apple Silicon, and there are ONNX INT8 builds that run on plain CPUs.

Grok Voice Transcribe 2.0 is a managed service, so the comparison is between a model and a product. What the product includes at its list price:

• Streaming — Grok Voice Transcribe 2.0 native over WebSocket, first partial at 0.49 seconds vs Parakeet TDT 0.6B's chunked or buffered approaches, with no first-class partial-transcript interface

• Key-term biasing — up to 100 key terms per request vs not a Parakeet feature

• Diarization — included in the request price vs a separate system you add yourself

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

• Inverse text normalization — included across 25 languages vs punctuation and capitalization only

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

• Licence — commercial API terms vs CC-BY-4.0 with attribution

• Model size — undisclosed vs about 600 million parameters, roughly 2.4 GB in fp32 or 1.2 GB in fp16

The last row is the one that decides a lot of deployments. Parakeet TDT 0.6B fits in a couple of gigabytes, runs acceptably on a laptop CPU through the INT8 ONNX path, and fits comfortably on any consumer GPU. That is not a smaller version of what the API does — it is a different capability, because it is the difference between a transcription feature that works offline, on a device, with no network and no per-hour cost, and one that does not.

The cost arithmetic nobody runs properly

The comparison that gets published is "$0.10 per hour versus free," and it is wrong in both directions — the API is not the only thing you pay for, and the weights are not the only thing you save on.

• Batch transcription — Grok Voice Transcribe 2.0 at $0.10 per audio hour, about $1.67 per 1,000 minutes vs Parakeet TDT 0.6B at no licence fee plus GPU or CPU time

• Streaming — $0.20 per audio hour, about $3.33 per 1,000 minutes vs self-hosted, where the constraint is your own concurrency ceiling rather than a published rate

• Service limits — 10 requests per second and 100 concurrent streaming sessions per team vs whatever your hardware allows, with no rate limit and no concurrency cap

• The cost that is not on either sheet — no engineering, no serving stack, no autoscaling vs the on-call rotation, the retry logic, the model updates and the GPU you keep warm for peak

At small volumes the managed side is almost always cheaper, because the self-hosted path's fixed cost is a person. At large steady volumes the self-hosted side wins decisively, and the crossover is a function of utilisation rather than audio volume: a GPU you keep busy is very cheap per hour of audio, and one you keep warm for peak traffic is not. A team processing 20,000 hours a month pays $2,000 for the managed batch path and roughly one mid-range GPU-month plus engineering time for the self-hosted one, which is not close.

The reason the arithmetic gets run badly is that the self-hosted side is usually compared at zero, and the managed side is usually compared at list. Neither is a real number. What is real is that Parakeet TDT 0.6B's 3,332 RTFx — NVIDIA's figure, batched, on datacenter hardware, and the one to treat as an upper bound rather than a promise — means a single modest GPU can process more audio than most organisations generate.

Where Parakeet stops being the right answer

Three things push a team off the open model and onto a managed one, and none of them is accuracy.

The first is key-term biasing. If your transcripts are full of product names, surnames, ticker symbols or domain jargon, a model with no biasing mechanism will get them wrong with no remedy short of fine-tuning. Grok Voice Transcribe 2.0 takes up to 100 key terms per request. Parakeet TDT 0.6B does not have the feature. For contact centres, healthcare and legal work, that single gap is disqualifying regardless of what any leaderboard says.

The second is diarization. Parakeet gives you words with timestamps; it does not tell you who said them. Multi-speaker transcription means running a separate diarization model and aligning the output, which is a project rather than a configuration option. The managed model returns speaker-attributed text in the same request.

The third is the streaming interface itself. Parakeet is fast enough that people build real-time systems on it — chunking audio, running the model on each chunk, stitching the output — but the partial-transcript behaviour, the end-of-turn detection and the commitment semantics are all things you write and maintain. Grok Voice Transcribe 2.0 returns a first partial 0.49 seconds after you stop speaking, as a product feature.

There is also a licensing detail that occasionally surprises teams: Parakeet TDT 0.6B V3 is CC-BY-4.0, which permits commercial use but requires attribution, and some NVIDIA speech models have since moved to the vendor's own Open Model License instead. If attribution is a problem for your product, read the card for the specific checkpoint rather than assuming the family's terms apply.

Why the API usually wins anyway, and when it should not

The pattern over the last year has been consistent: teams start on an open model like Parakeet TDT 0.6B because it is free and controllable, ship the feature, then discover that the ongoing cost is not the GPU but the maintenance, and migrate to a managed endpoint. The reverse migration happens too, usually when volume crosses a threshold where the per-hour charge starts to look like rent on something you could own.

Both directions are expensive to execute if the model is wired directly into your application, which is the argument for keeping the call site boring. OrcaRouter exposes 200+ models behind a single OpenAI-compatible key with automatic failover across providers and 0% markup on provider list price, so a self-hosted deployment and a managed one can sit behind the same interface and be swapped with a model string. That matters more than usual in speech, where the leaderboard changes hands every few weeks and a single-region managed service is a single-region outage — the failover path is what keeps a transcription feature from being a transcription outage.

For teams that want the throughput of the open model without the serving stack, that is also the shape of the decision: benchmark Parakeet TDT 0.6B on your own audio, benchmark Grok Voice Transcribe 2.0 on the same audio, and let the one that wins keep the traffic while you stop caring which vendor's logo is on it.

If you want the one-line version: Parakeet TDT 0.6B is still the best answer to "transcribe anything, anywhere, at no per-hour cost, on hardware I already own," and Grok Voice Transcribe 2.0 is the answer to "transcribe accurately with speaker labels and my own vocabulary, without running anything." The 13.7% figure that makes the gap look enormous is a retired measurement, and the honest gap — somewhere between one and three points of WER on overlapping audio — is small enough that the operational question should decide it.