faster-whisper is a reimplementation of Whisper using CTranslate2, an inference engine for Transformer models. It supports quantization on CPU and GPU and reports faster transcription with lower memory use than the reference implementation. Its benchmark results depend on model, precision, beam size and batching, so those settings are essential to understanding the reported transcription speed.
| Model / device | Implementation | Precision | Beam size | Time | Memory |
|---|---|---|---|---|---|
| Large-v2 / GPU | Whisper reference | fp16 | 5 | 2m23s | 4708MB VRAM |
| Large-v2 / GPU | faster-whisper | fp16 | 5 | 1m03s | 4525MB VRAM |
| Large-v2 / GPU | faster-whisper (batch_size=8) | int8 | 5 | 16s | 4500MB VRAM |
| Small / CPU | Whisper reference | fp32 | 5 | 6m58s | 2335MB RAM |
| Small / CPU | faster-whisper | fp32 | 5 | 2m37s | 2257MB RAM |
| Small / CPU | faster-whisper (int8) | int8 | 5 | 1m42s | 1477MB RAM |
Reported README measurements. The GPU benchmark transcribes 13-minute audio. CPU and GPU rows use different models; batched and unbatched rows also use different settings. These are not WhisperJot benchmarks.
The README claims up to 4 times faster transcription than the Whisper reference implementation at the same accuracy while using less memory. It also supports 8-bit quantization on CPU and GPU. Read that claim alongside the table: changing batching and precision changes the experiment. Batched throughput measures processing across a batch; the delay after a single dictated utterance is a separate measurement.
faster-whisper requires Python 3.9 or greater. The documented GPU path requires cuBLAS for CUDA 12 and cuDNN 9 for CUDA 12. Example compute types include float16, int8 and int8_float16. This makes it a candidate when you want a Python transcription integration with explicit precision choices. Verify your environment against the runtime requirements before selecting a GPU configuration for deployment.
The runtime is MIT licensed and supports Distil-Whisper checkpoints. Its README specifically says distil-large-v3 was designed to work with the faster-whisper transcription algorithm. Consider it when your application already needs Python and you want to evaluate CPU quantization or batched processing. Keep the checkpoint and workload fixed when comparing implementations, and measure both transcript quality and resource use on representative audio.
WhisperJot does not use faster-whisper. Its Jot Local Pro engine uses WhisperKit on macOS and whisper.cpp-format GGML models through a .NET runtime on Windows and Linux. Jot Local uses Parakeet instead. The faster-whisper benchmarks measure a separate runtime. WhisperJot's performance depends on its own engines and model selections. Auto mode chooses between the product's local engines per utterance; both work offline after their model downloads.
Explore WhisperJot →It is a Whisper reimplementation built on CTranslate2. The runtime affects how the checkpoint executes, including precision and batching options. You still choose the model. Distil-Whisper checkpoints are also supported, but changing the checkpoint changes more than the runtime. Keep both choices explicit.
Yes. The README provides CPU benchmarks and supports 8-bit quantization on CPU as well as GPU. Its Small-model CPU table includes fp32 and int8 results. Those measurements describe the documented benchmark, not a fixed speed on every computer. Evaluate your own recordings too.
No. WhisperJot uses WhisperKit for macOS Jot Local Pro and whisper.cpp-format GGML models on Windows and Linux. Its other local engine is Parakeet-based. The faster-whisper performance figures measure that runtime alone; WhisperJot uses different runtimes. Choose the runtime independently when building a Python application.
Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.
Explore whisper.cpp, its GGML model disk sizes, integer quantization and Apple acceleration options, plus how WhisperJot uses GGML on Windows and Linux.
Understand Distil-Whisper's English-only student models, decoder changes and reported speed, size and WER claims, including the distil-large-v3 checkpoint.
Plan local Whisper transcription with the official VRAM table, whisper.cpp and Python routes, or WhisperJot's desktop models and honest download sizes.
See also: Whisper glossary definition · System requirements · Pricing
