Parakeet TDT 0.6B v2, WER 6.05%
Vast.ai RTX 4090 24GB on-demand median
Services
Transcribing 1,000 hours of English audio a month as a batch job, Canary-Qwen 2.5B (5.63% WER on the Open ASR Leaderboard) on a Vast.ai RTX 4090 needs about 4.8 GPU-hours, $1.96 in rental, against $258 for Deepgram Nova-3 and $180 for OpenAI gpt-4o-mini-transcribe. The GPU price is a marketplace median. The speed is an estimate.
Enter your own audio hours below. The calculator picks an open model and a GPU, then compares that bill with OpenAI, Deepgram, AssemblyAI, Google, AWS, and hosted Whisper.
$1.96
this month · $0.002 per audio hour · or $295.24 if the GPU stays on
The pick is the lowest WER among setups whose bill is no more than twice the cheapest one, or $10 more, whichever is larger.
WER is the average word error rate on the Open ASR Leaderboard English short-form sets (AMI, Earnings-22, GigaSpeech, LibriSpeech, SPGISpeech, TED-LIUM, VoxPopuli). Lower is better.
Parakeet TDT 0.6B v2, WER 6.05%
Vast.ai RTX 4090 24GB on-demand median
Parakeet TDT 0.6B v2 on Vast.ai A100 SXM4 80GB on-demand median. RTFx is the board’s A100 80GB SXM measurement.
Free, from this page
One email with this page’s inputs, the recommended model, GPU, and provider, a deployment checklist, and whether self-hosting or the API is the smaller bill.
Speech-to-text deployment plan
Canary-Qwen 2.5BVast.ai · NVIDIA GeForce RTX 4090
$1.96 this month
Deepgram $258.00 for the same work. GPU rental only.
At 1,000 audio hours a month in batch, Canary-Qwen 2.5B on a Vast.ai RTX 4090 is $1.96 of GPU rental. Deepgram Nova-3 is $258, OpenAI gpt-4o-mini-transcribe $180, and Amazon Transcribe $360. The cheapest hosted open model is Groq whisper-large-v3-turbo at $40.
The batch number assumes you only pay for GPU hours you use. Left on for all 730 hours, the same GPU is $295.24 a month. That is where most self-hosted speech projects lose money: a GPU sized for peak streaming traffic sits idle overnight. Both figures are GPU rental only, with no engineering, storage, or egress.
WER is the average word error rate on the Open ASR Leaderboard English short-form sets (AMI, Earnings-22, GigaSpeech, LibriSpeech, SPGISpeech, TED-LIUM, VoxPopuli). Lower is better. Every score links to the leaderboard. Pick a model to see every GPU and provider for it.
| Model | WER | Languages | GPU | Provider | $/hour | Always-on / month | This workload | Per audio hour | RTFx per GPU |
|---|---|---|---|---|---|---|---|---|---|
| Parakeet TDT 0.6B v2 CC-BY-4.0 | 6.05% | English | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $0.24 | $0.0002 | 1,674 estimate |
| Parakeet TDT 0.6B v3 CC-BY-4.0 | 6.32% | 25 European languages | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $0.25 | $0.0002 | 1,648 estimate |
| Canary 1B Flash CC-BY-4.0 | 6.35% | English, German, Spanish, French | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $0.78 | $0.0008 | 517 estimate |
| Canary-Qwen 2.5B · recommended CC-BY-4.0 | 5.63% | English | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $1.96 | $0.002 | 207 estimate |
| Qwen3-ASR 1.7B Apache-2.0 | 5.76% | Not in the card metadata | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $5.53 | $0.0055 | 73 estimate |
| Granite Speech 3.3 8B Apache-2.0 | 5.74% | English, French, German, Spanish, Portuguese | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $5.63 | $0.0056 | 72 estimate |
| Whisper large-v3 Apache-2.0 | 7.44% | 99 languages | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $5.62 | $0.0056 | 72 estimate |
| Whisper large-v3-turbo MIT | 7.83% | 99 languages | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $4.09 | $0.0041 | 99 estimate |
| Distil-Whisper large-v3.5 MIT | 7.21% | English | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $4.05 | $0.004 | 100 estimate |
| Voxtral Mini 3B Apache-2.0 | 7.05% | 8 languages in the card | NVIDIA GeForce RTX 4090 | Vast.ai RTX 4090 24GB on-demand median | $0.4044 median | $295.24 | $7.45 | $0.0074 | 54 estimate |
| Voxtral Small 24B Apache-2.0 | 6.62% | 8 languages in the card | NVIDIA GeForce RTX 4090 × 4 | Vast.ai RTX 4090 24GB on-demand median | $1.6178 median | $1,181 | $15.13 | $0.0151 | 27 estimate |
faster-whisper (CTranslate2) is a runtime for the Whisper weights, not a separate model. It has no row on the leaderboard, so it gets no WER or RTFx of its own here. Expect the Whisper large-v3 accuracy and a different speed. “Measured” means the board’s A100 80GB SXM figure. Every other GPU is a bandwidth estimate.
List prices checked November 1, 2026. A dash in WER means that API has no row on the Open ASR Leaderboard table used here. A dash in a price means that API has no such mode.
| API | Type | Batch $/min | Real-time $/min | This month | Per audio hour | WER |
|---|---|---|---|---|---|---|
| OpenAI gpt-4o-mini-transcribe | Managed API | $0.003 | — | $180.00 | $0.18 | — |
| OpenAI gpt-transcribe | Managed API | $0.0045 | — | $270.00 | $0.27 | — |
| OpenAI gpt-realtime-whisper | Managed API | — | $0.017 | — | — | — |
| Deepgram Nova-3 (monolingual) | Managed API | $0.0043 | $0.0077 | $258.00 | $0.26 | — |
| AssemblyAI Universal-3.5 Pro / 3.6 Pro Realtime | Managed API | $0.0035 | $0.0075 | $210.00 | $0.21 | 6.21% |
| AssemblyAI Universal-2 / Universal-Streaming | Managed API | $0.0025 | $0.0025 | $150.00 | $0.15 | — |
| Amazon Transcribe | Managed API | $0.006 | $0.01 | $360.00 | $0.36 | — |
| Google Cloud STT v2 | Managed API | $0.003 | tiered | $180.00 | $0.18 | 6.63% |
| Groq whisper-large-v3-turbo | Hosted open model | $0.00067 | — | $40.00 | $0.04 | 7.83% |
| Groq whisper-large-v3 | Hosted open model | $0.00185 | — | $111.00 | $0.11 | 7.44% |
| Together Whisper large-v3 | Hosted open model | $0.0015 | $0.0035 | $90.00 | $0.09 | 7.44% |
| Together Parakeet TDT 0.6B v3 | Hosted open model | $0.0015 | — | $90.00 | $0.09 | 6.32% |
| DeepInfra Voxtral Mini | Hosted open model | $0.001 | — | $60.00 | $0.06 | 7.05% |
AssemblyAI’s 6.21% is the board row universal-3-pro, and Google’s 6.63% is chirp_3. Hosted open models show the WER of the open weights they serve.
On-demand hourly price and the always-on month (hourly × 730). Community and non-preemptible rates are shown only where that provider publishes them. Each price links to the provider’s pricing page, checked November 1, 2026. A dash means that provider does not list this GPU.
| GPU | RunPod | Lambda | Modal | Vast.ai | AWS | Google Cloud |
|---|---|---|---|---|---|---|
| L4 | — | Nvidia L4 $0.7992/hr Preemptible $583.42/mo Non-preemptible $2.3976/hr · $1,750/mo checked November 1, 2026 | ||||
| A10 / A10G | — | Nvidia A10 $1.1016/hr Preemptible $804.17/mo Non-preemptible $3.3048/hr · $2,413/mo checked November 1, 2026 | — | |||
| RTX 4090 | — | — | — | — | ||
| A100 | Nvidia A100 80GB $2.4984/hr Preemptible $1,824/mo Non-preemptible $7.4952/hr · $5,471/mo checked November 1, 2026 | |||||
| H100 | Nvidia H100 SXM5 $3.9492/hr Preemptible $2,883/mo Non-preemptible $11.85/hr · $8,649/mo checked November 1, 2026 |
Prices checked November 1, 2026. A month is 730 hours. The bill is GPU rental only. WER is the average word error rate on the Open ASR Leaderboard English short-form sets (AMI, Earnings-22, GigaSpeech, LibriSpeech, SPGISpeech, TED-LIUM, VoxPopuli). Lower is better. The pick is the lowest WER among setups whose bill is no more than twice the cheapest one, or $10 more, whichever is larger.
RTFx is audio seconds transcribed per second of GPU time. The board measured it on one A100-SXM4-80GB at batch size 64. Other GPUs use that RTFx × (the GPU's NVIDIA memory bandwidth ÷ 2,039 GB/s), and are marked estimates. A model that needs more than one GPU is assumed to scale linearly with the GPU count. A 24 GB card may not hold batch 64 for the larger models, so its scaled rate can run high. RTFx is a batch number. A streaming server that holds latency down serves fewer streams than RTFx, so the real-time GPU count here is a floor.
Related tools: the self-hosted OCR cost calculator, the embedding cost calculator, and the open-source LLM hosting cost calculator. Related reading: LLM API pricing, LLM cost optimization, and AI agent development.
Usually, for batch audio. OpenAI gpt-4o-mini-transcribe is $0.003 a minute, $0.18 per audio hour. The leaderboard measured Whisper large-v3 at an RTFx of 145.5 on one A100 80GB SXM, so an hour of audio takes under 25 seconds of that GPU. Even at RunPod's $1.79 an hour that is about $0.01 of GPU time per audio hour. The catch is utilization: a GPU you leave on all month costs the same whether you send it audio or not.
On the Open ASR Leaderboard English short-form table (commit of 22 May 2026), the lowest average WER among the open models here is NVIDIA Canary-Qwen 2.5B at 5.63, then Granite Speech 3.3 8B at 5.74 and Qwen3-ASR 1.7B at 5.76. Whisper large-v3 is 7.44. These are English scores. Check the per-language results before using any of them for other languages.
Inverse real-time factor: seconds of audio transcribed per second of compute. An RTFx of 3,386, which the board reports for Parakeet TDT 0.6B v2, means one GPU-second transcribes about 56 minutes of audio. The board measured RTFx on one A100-SXM4-80GB at batch size 64. On other GPUs this page scales it by memory bandwidth and labels the result an estimate.
They have no row on the Open ASR Leaderboard table used here. The page shows a dash instead of borrowing a number from a vendor's own benchmark. AssemblyAI and Google do have board rows, but under model names that are not exactly the SKU on the price page, and the table says which row was used.
This calculator divides your peak concurrent streams by the batch RTFx of the setup and rounds up, with at least one GPU on all month. That is a floor. A streaming server that keeps latency low runs smaller batches than the benchmark, so it serves fewer streams per GPU. Load-test your own server with your stream count before you size it.
Not as its own row. faster-whisper is a CTranslate2 runtime for the Whisper weights, so its accuracy is the Whisper large-v3 accuracy and its speed depends on the GPU, batch size, and quantization. This page does not give it a separate RTFx.