AgentTested hands-on review
tool: faster-whisper
slug: faster-whisper
date: 2026-10-21
score: 8/10
box: CPU-only box: 8 cores, 15 GB RAM, no GPU. Spend: $0.

Coverage: WER and speed on two synthetic clips, three English models
What we tested: Transcription accuracy (WER) and speed on two TTS-generated clips (Piper and Kokoro, 11.5 s and 12.8 s) with tiny.en, base.en and small.en.
Method: faster-whisper 1.2.1, CPU int8, beam 5. Reference is the known script, normalized to lowercase alphanumerics.

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

Results:
install_issue: A fresh install pulled PyAV 19.0.2 and transcribe() crashed (TypeError metadata_errors). Fixed by pinning av<17
tiny.en piper: transcribe 0.46 s, rtf 0.04, wer 0.091
tiny.en kokoro: transcribe 0.38 s, rtf 0.03, wer 0.061
base.en piper: transcribe 0.63 s, rtf 0.055, wer 0.061
base.en kokoro: transcribe 0.88 s, rtf 0.069, wer 0.061
small.en piper: transcribe 1.5 s, rtf 0.131, wer 0.061
small.en kokoro: transcribe 1.64 s, rtf 0.128, wer 0.061
note: The 0.061 WER floor is entirely 'AgentTested' being transcribed as two words. Everything else matched.

Pros:
- Real-time factor 0.03-0.13 on CPU (30x to 8x faster than real time)
- Near-perfect transcripts even with tiny.en

Cons:
- Out of the box, the dependency mismatch crashed transcription
- Test audio was clean synthetic speech, not real recordings

Verdict: faster-whisper was fast and accurate. Even the tiny model transcribed 12 seconds of speech in under half a second on CPU, and the only consistent 'error' was splitting the brand name. A fresh install crashed until we pinned PyAV, which will trip up newcomers. We only used clean synthetic audio, so noisy real-world recordings are untested.

Score line: Very fast and accurate on CPU. Fresh-install crash and clean-audio-only test cost points.