faster-whisper vs whisper.cpp
The same Whisper weights in two runtimes. On our x86 CPU box faster-whisper (int8) was 1.6–2.9x faster, while whisper.cpp (fp16) used about 20–25% less RAM and loaded faster. The faster-whisper project's own i7 benchmark agrees that int8 wins, but there whisper.cpp beats faster-whisper when both run fp32. On Apple Silicon only whisper.cpp uses the GPU (Metal/Core ML).
Measured by us: same weights, two runtimes
| Model | Runtime / settings | RTF | First segment | Peak RAM | Load | WER* |
|---|---|---|---|---|---|---|
| Whisper tiny | faster-whisper tiny (int8, beam 5) | 0.0291 | 0.294 s | 290.4 MB | 0.46 s | 9.7% |
| Whisper tiny | whisper.cpp tiny (ggml fp16, beam 5, 8 threads, pywhispercpp) | 0.0457 | 0.461 s | 226.4 MB | 0.08 s | 9.7% |
| Whisper base | faster-whisper base (int8, beam 5) | 0.0410 | 0.413 s | 443.0 MB | 3.64 s | 6.5% |
| Whisper base | whisper.cpp base (ggml fp16, beam 5, 8 threads, pywhispercpp) | 0.0786 | 0.793 s | 334.7 MB | 0.26 s | 3.2% |
| Whisper small | faster-whisper small (int8, beam 5) | 0.1008 | 1.017 s | 989.8 MB | 4.11 s | 0.0% |
| Whisper small | whisper.cpp small (ggml fp16, beam 5, 8 threads, pywhispercpp) | 0.2958 | 2.984 s | 800.9 MB | 0.78 s | 0.0% |
Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU, 2026-10-21, beam 5 for both, median of 5 runs, 10.1 s clean synthetic English clip. *WER is a sanity check only. Methodology
Side by side
| Whisper tiny | Whisper base | Whisper small | |
|---|---|---|---|
| Organization | OpenAI | OpenAI | OpenAI |
| Parameters | 38M | 73M | 242M |
| License | apache-2.0 source | apache-2.0 source | apache-2.0 source |
| Commercial use | Yes Apache-2.0 on HF card; MIT on GitHub | Yes Apache-2.0 on HF card; MIT on GitHub | Yes Apache-2.0 on HF card; MIT on GitHub |
| Stated memory | ~1 GB VRAM | ~1 GB VRAM | ~2 GB VRAM |
| Runs on CPU | Yes | Yes | Yes |
| Streaming | No | No | No |
| Languages | 99 languages | 99 languages | 99 languages |
| Our fastest CPU result | RTF 0.0291 faster-whisper tiny (int8, beam 5); Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU | RTF 0.0410 faster-whisper base (int8, beam 5); Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU | RTF 0.1008 faster-whisper small (int8, beam 5); Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU |
| Pi 5 | Sourced statement | Sourced statement | CPU supported, no device data |
| CPU only | Measured by us | Measured by us | Measured by us |
| Apple Silicon | Fits stated memory | Fits stated memory | Sourced statement |
| 8 GB GPU | Fits stated memory | Fits stated memory | Fits stated memory |
| 12–16 GB GPU | Fits stated memory | Fits stated memory | Fits stated memory |
| 24 GB+ GPU | Fits stated memory | Fits stated memory | Fits stated memory |
Each fact is sourced on the model's page. "Unknown" means the source doesn't say. Accuracy: Hugging Face Open ASR Leaderboard.
Key differences (sourced)
- faster-whisper README (small, CPU, i7-12700K, 8 threads, beam 5): whisper.cpp fp32 2m05s / 1049 MB vs faster-whisper fp32 2m37s / 2257 MB and int8 1m42s / 1477 MB. source
- faster-whisper README (large-v2, RTX 3070 Ti 8 GB, beam 5): whisper.cpp with flash attention 1m05s / 4127 MB vs faster-whisper fp16 1m03s / 4525 MB and int8 59s / 2926 MB. source
- whisper.cpp supports Apple Metal, Core ML (ANE), CUDA, Vulkan, OpenVINO and plain C/C++ CPU builds with no Python. source
- Our whisper.cpp numbers use the prebuilt pywhispercpp 1.5.1 wheel (AVX2, no AVX-512). A native build tuned for this CPU could be faster. source
Which should you pick?
- Pick faster-whisper on x86 or NVIDIA in Python, especially with int8 or batched inference.
- Pick whisper.cpp on Macs, Raspberry Pi, or when you need a dependency-free C/C++ binary or the lowest RAM use.
Reach developers picking a speech model to self-host. Advertise here →
Sponsors never affect rankings, picker results or benchmark numbers.