ANVE: the private AI office, on the record. See what that means
ANVEPrivate AI, answered · updated 2026-10-16
Short answer. From about one minute to half an hour, depending on the program, the model size and the hardware. Scaling the faster-whisper project’s 13-minute benchmark to 60 minutes, the large-v2 model on an RTX 3070 Ti would take about 1.2 to 11 minutes, and the small model on an 8-thread Core i7-12700K about 3.9 to 32 minutes. These are estimates by simple proportion from one set of recordings, not promises. Hindi, noisy rooms and several speakers change the result.
The faster-whisper project publishes the time and memory needed to transcribe 13 minutes of audio with several programs. This page multiplies each time by 60 ÷ 13, about 4.6, to show an hour. That assumes the time grows in proportion to the audio, which is a fair approximation for batch transcription of similar recordings and is not a measurement. Every figure here is an estimate.
The benchmark ran on an NVIDIA RTX 3070 Ti with 8 GB, using CUDA 12.4. The large-v2 model has the same size as large-v3.
| Program and precision (beam size 5) | Measured, 13 minutes | Estimated, 60 minutes |
|---|---|---|
| openai/whisper, 16-bit | 2 min 23 s | 11.0 min |
| transformers (SDPA), 16-bit | 1 min 52 s | 8.6 min |
| whisper.cpp with flash attention, 16-bit | 1 min 05 s | 5.0 min |
| faster-whisper, 16-bit | 1 min 03 s | 4.8 min |
| faster-whisper, 8-bit | 59 s | 4.5 min |
| faster-whisper, 16-bit, batch size 8 | 17 s | 1.3 min |
| faster-whisper, 8-bit, batch size 8 | 16 s | 1.2 min |
The CPU runs used 8 threads on an Intel Core i7-12700K and the small model, which is less accurate than large.
| Program and precision (beam size 5) | Measured, 13 minutes | Estimated, 60 minutes |
|---|---|---|
| openai/whisper, 32-bit | 6 min 58 s | 32.2 min |
| whisper.cpp, 32-bit | 2 min 05 s | 9.6 min |
| whisper.cpp with OpenVINO, 32-bit | 1 min 45 s | 8.1 min |
| faster-whisper, 32-bit | 2 min 37 s | 12.1 min |
| faster-whisper, 32-bit, batch size 8 | 1 min 06 s | 5.1 min |
| faster-whisper, 8-bit | 1 min 42 s | 7.8 min |
| faster-whisper, 8-bit, batch size 8 | 51 s | 3.9 min |
Every row above finishes faster than the recording lasts, the slowest being a 6 min 58 s run for 13 minutes of audio, which is 1.9 times real time. Live captioning is a streaming problem that these batch figures do not measure. ANVE’s meeting minutes start from a transcript in Hindi, English or Hinglish; ANVE has not published timings for its launch hardware.
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.