Speech-to-text Measured by us

Whisper small on CPU only (Intel N100 mini-PC / x86 desktop)

We ran it. Results below, with the exact hardware and settings.

No GPU. Covers low-power N100 boxes up to desktop CPUs. Our own CPU numbers come from an 8 vCPU cloud Xeon, which is faster than an N100.

EvidenceMeasured by us

Commercial useYes

Licenseapache-2.0

Stated memory~2 GB VRAM

Device memory16 GB RAM

Results

Runtime / settings	RTF	First-segment latency	Peak RAM	Load	WER*
faster-whisper small (int8, beam 5)	0.1008 (10× real time)	1.017 s	989.8 MB	4.11 s	0.0%
whisper.cpp small (ggml fp16, beam 5, 8 threads, pywhispercpp)	0.2958 (3× real time)	2.984 s	800.9 MB	0.78 s	0.0%

Measured on Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU (Intel(R) Xeon(R) Processor, 8 threads, Linux 6.12.94+, Python 3.13.5) on 2026-10-21. This is a cloud Xeon, not an N100: expect an N100 (4 cores, 6 W) to be several times slower. We'll post real N100 numbers when we have one. *WER on clean synthetic speech is a sanity check, not real-world accuracy.

What the sources say about CPU only

Third-party docs faster-whisper README, small model on CPU (Intel Core i7-12700K, 8 threads, beam 5): whisper.cpp fp32 2m05s / 1049 MB; faster-whisper fp32 2m37s / 2257 MB; faster-whisper int8 1m42s / 1477 MB; int8 batch 8: 51s / 3608 MB. source
Caveat: Benchmark published by the faster-whisper project; audio length per their README

Method

Harness: bench/run.py (open source), 5 timed runs after 1 warm-up, median reported.
Input: a 10.1 s English clip synthesized with Piper (en_US-lessac-medium), resampled to 16 kHz mono, language forced to English. First-segment latency = time until the first transcribed segment.
RTF = wall-clock processing time ÷ audio duration (lower is faster). Model load time is reported separately, not included.
Peak RAM = whole-process RSS sampled every 20 ms, including model weights.
Machine: Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU. No other benchmark ran at the same time; a browser was open on the same VM.

Ways to run Whisper small

Runtime	Precision	Documented devices
openai-whisper (PyTorch)	fp16/fp32	CPU, CUDA	source
faster-whisper (CTranslate2)	int8, fp16, fp32	CPU, CUDA	source
whisper.cpp (ggml)	fp16, quantized (q5/q8)	CPU (x86/ARM), Apple Metal + Core ML, CUDA, Vulkan, OpenVINO	source
Hugging Face transformers	fp16/fp32	CPU, CUDA	source

Run it

pip install faster-whisper "av<16"   # av 19 broke file decoding with faster-whisper 1.2.1 when we tested
python -c "from faster_whisper import WhisperModel; m=WhisperModel('small', device='cpu', compute_type='int8'); segs,_=m.transcribe('audio.wav', beam_size=5); print(' '.join(s.text for s in segs))"

We ran this exact command on our test box on 2026-10-21.

Caveats

Our CPU numbers come from an 8 vCPU cloud Xeon, which is several times faster than an Intel N100
Benchmark published by the faster-whisper project; audio length per their README