AgentTested hands-on review
tool: whisper.cpp
slug: whisper-cpp
date: 2026-10-21
score: 8/10

Coverage: Build, plus the same Piper and Kokoro clips used for faster-whisper
What we tested: Build friction and transcription of the same Piper and Kokoro clips used for faster-whisper.
Method: cmake build (Release), ggml-base.en via the repo's download script, whisper-cli -t 8 on 16 kHz mono conversions.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- Synthetic audio only

Results:
build_s: 77
model_download_s: 2
piper: 1.75 s wall incl. load for 11.5 s audio; transcript matches script except 'AgentTested' split (WER 0.061)
kokoro: 1.68 s wall incl. load for 12.8 s audio; same WER 0.061

Pros:
- No Python and no dependency mismatch issues
- Accurate; about 7x faster than real time including model load

Cons:
- Needs a compile and 16 kHz WAV input (we converted with ffmpeg)

Verdict: whisper.cpp matched faster-whisper's accuracy on our clips and ran in under 2 seconds including model load, with none of the Python dependency trouble. You compile it yourself and feed it 16 kHz audio, but after that it just works.

Score line: Accurate and fast with no dependency trouble. Synthetic audio only.