How to run Whisper voice models on a refurbished PC

Whisper Large v3 is the model most transcription pipelines settle on. It handles accents, background noise, and long recordings better than the smaller variants, and the licence permits commercial use. The catch is that OpenAI's reference implementation was written on hardware most small operators do not own. A refurbished business workstation, bought sensibly, closes almost all of that gap for a fraction of the price.

Below: what runs on CPU, what a modest GPU changes, how much RAM is really needed, and why a refurbished HP EliteBook makes a strong voice pipeline.

What Whisper Large v3 asks of a machine

The Large v3 checkpoint is roughly 1.55 billion parameters, a 3.1 GB file in float16. Loaded in native precision it wants around 6 GB of RAM before a single second of audio has been processed. At int8 quantisation, which is what most self hosters use, that drops to about 3.2 GB. Add the operating system, the Python runtime, and an audio buffer, and 16 GB of system RAM is the sensible floor. 32 GB adds headroom for batching, or for a second model such as a diarisation pass in the same process.

Whisper also wants single thread performance. The encoder does a lot of matrix multiplication that benefits from AVX2 and, on newer chips, AVX512. Any Intel processor from the eighth generation onwards, or any Ryzen from the 3000 series onwards, does the job. Older Xeons technically work but feel sluggish.

CPU only: what real time actually means

The industry shorthand is the real time factor, or RTF. An RTF of 1.0 means one second of audio takes one second to transcribe. Below 1.0 is real time. Above 1.0 is batch.

On an Intel Core i7 of the tenth or eleventh generation, running Whisper Large v3 through faster-whisper at int8, expect an RTF of roughly 0.6 to 0.9 for clean sixteen kilohertz audio. That is real time with margin. Push the same workload through the original openai-whisper package and the RTF climbs to 2.0 to 3.0, because that codebase does not use the CTranslate2 backend. The library choice matters more than most people realise.

For a call centre doing overnight batch transcription, or a solicitor's office turning around dictations by end of day, CPU only is entirely adequate. A tenth generation i7 will churn through an hour of audio in about forty to fifty minutes.

See the BAV HP EliteBook listing on eBay

When a GPU earns its money

For genuine live transcription with sub second latency, a GPU changes the picture. A used NVIDIA RTX 3060 with 12 GB of VRAM runs Whisper Large v3 in float16 at an RTF of roughly 0.05 to 0.08. That is twenty times faster than the audio itself. A single card handles four to six concurrent live streams before latency creeps up.

For media monitoring, live captioning, or a voice agent that must respond inside a second, the GPU is not optional. For everything else, it is nice to have. A used 3060 for around 180 pounds is one of the best value upgrades in the current market.

Storage and the boring bits that matter

Model files load from disk into RAM at startup. On a SATA SSD, cold start for Whisper Large v3 takes about four seconds. On NVMe it is roughly one and a half. If workers restart frequently, NVMe pays for itself. Whisper resamples audio to sixteen kilohertz mono internally, so a two hour recording is under 240 MB even in uncompressed WAV.

Why refurbished workstations suit voice pipelines

Voice transcription rewards steady, sustained performance rather than peak bursts. That is what business class refurbished hardware was engineered for. An HP EliteBook, a Dell Latitude, or a Lenovo ThinkPad from the tenth generation Intel era was specified for eight hour days of continuous office work, with cooling and power delivery to match. They throttle less under sustained load than consumer laptops of the same vintage.

The BAV HP EliteBook range covers a spread of configurations: Intel Core i5 or i7 of the eighth to eleventh generation, 8 GB to 32 GB of DDR4 RAM, and 256 GB to 1 TB of NVMe or SATA SSD storage. For a Whisper pipeline the sweet spot is the i7 with 16 GB or 32 GB of RAM and a 512 GB NVMe.

Putting it together

A practical starting point: install Ubuntu 24.04 LTS or Windows 11 Pro. Install Python 3.11 and create a virtual environment. Install faster-whisper and its CTranslate2 dependency. Download the Large v3 model in int8 precision, a one line command through the library. Wrap the transcription call in a small FastAPI service on a local port and point the existing audio source at that endpoint. The stack fits inside 4 GB of RAM at idle and scales linearly with concurrent requests to the limits of the CPU.

FAQ

Can Whisper Large v3 run on 8 GB of RAM?

Technically yes, at int8 quantisation, but there is almost no headroom for the operating system or other processes. Expect swap thrashing on longer files. 16 GB is the realistic minimum for reliable operation.

Is the smaller Whisper model good enough?

For clear English audio recorded on a decent microphone, the medium model gets within a percentage point or two of Large v3 on word error rate and runs roughly twice as fast. For accented English, noisy calls, or non English languages, the gap widens quickly and Large v3 is worth the extra compute.

Does it work on AMD Ryzen?

Yes. Ryzen 3000 series and newer handle Whisper on CPU comfortably, and often outperform Intel chips of the same era on multi threaded workloads. The faster-whisper library is CPU vendor agnostic. Just make sure the machine has AVX2 support, which every Ryzen since 2017 does.

What about Apple Silicon?

Whisper runs very well on M1 and M2 chips through the mlx-whisper port, with RTF figures roughly comparable to a mid range NVIDIA card. If one is already owned, it is a strong option. For a fresh purchase, a refurbished Intel or AMD workstation gives more RAM and storage for the same money.

About Birmingham AV

Birmingham AV has sold over 87,000 items on eBay since 2017 and holds 24,756 buyer feedbacks at 98.9% positive. Every machine we ship carries a twelve month warranty and is tested before dispatch from our facility in Bromsgrove, Worcestershire. We are Companies House 12383651 and VAT registered as GB 348755066. We are one of the highest volume refurbished PC operations on eBay UK, and every unit is prepared, photographed, and dispatched by the same team that answers your questions.