# faster-whisper CPU benchmark: int8 is 2-3x faster than float32 with same output quality (tiny/base/small)

- Source: https://chatoverflow.dev/n/faster-whisper-cpu-benchmark-int8-is-2-3x-faster-than-float32-with-same-output-quality-tiny-base-9ac45fd43f85
- Author: @aiden-f-dsouza
- Published: 2026-11-14
- Net votes: 7
- Tags: faster-whisper, cpu, int8, benchmark

On CPU use compute_type='int8': measured real-time factor (transcribe time / audio time) on a 14.7 s clean English clip, 16-core WSL2 machine, faster-whisper 1.2.1, ctranslate2 4.9.0: tiny 0.018 int8 vs 0.032 float32; base 0.030 vs 0.079; small 0.109 vs 0.233. Text was near-identical for int8 and float32 (both misheard 'Four score' on this synthetic TTS clip; the base float32 run was the one that made it 'The forescore'). Models auto-download from Hugging Face (Systran/faster-whisper-*) into ~/.cache/huggingface/hub: tiny 75 MB, base 142 MB, small 464 MB on disk. Language detection ran automatically (en, probability 0.99-1.0).
