AI Productivity & Automationproductivitylocal aibeginners

faster-whisper vs whisper.cpp vs Buzz: Which Is Fastest?

If Whisper feels slow you are using the wrong implementation. Real benchmarks across four options.

Published October 3, 2026

Muhammad Usman

By Muhammad Usman · Founder & Lead Reviewer

faster-whisper vs whisper.cpp vs Buzz: Which Is Fastest?

Quick Answer

faster-whisper transcribes 13 minutes of audio in 59 seconds using 2.9GB of memory. The original openai/whisper takes 2 minutes 23 seconds and 4.7GB for the same job with the same model. If Whisper feels slow, you are almost certainly using the wrong implementation.

faster-whisper transcribes 13 minutes of audio in 59 seconds using 2.9GB of memory. The original openai/whisper takes 2 minutes 23 seconds and 4.7GB for the same job with the same model. If Whisper feels slow, you are almost certainly using the wrong implementation.

Key takeaways

  • faster-whisper at int8 is 2.4x faster on 40% less memory than openai/whisper, with identical output quality.
  • whisper.cpp is the CPU-only choice, running on anything from a Mac to a Raspberry Pi with no dependencies.
  • Buzz is the beginner choice: a real desktop app with installers, speaker identification and subtitle export.
  • All four are MIT licensed, so transcripts carry no commercial restrictions.
  • openai/whisper does not do speaker diarization. Buzz and WhisperX do.

The benchmark

From faster-whisper's own README, using Whisper large-v2 on 13 minutes of audio with an NVIDIA RTX 3070 Ti (8GB):

ImplementationPrecisionTimeVRAM
openai/whisperfp162m23s4,708MB
whisper.cpp (Flash Attention)fp161m05s4,127MB
faster-whisperfp161m03s4,525MB
faster-whisperint859s2,926MB

A CPU-only datapoint from the same source: the small model at fp32 on an Intel Core i7-12700K took 2m37s with faster-whisper.

These are the same model weights producing the same transcript. The difference is entirely in how the inference is implemented.

What each one is for

openai/whisper: the reference

The original implementation from OpenAI. Python, straightforward, and the definitive version of the model. 108,292 GitHub stars, MIT licensed for both code and weights.

Use it if you want the canonical behaviour or are experimenting. For production transcription, the alternatives below are simply faster.

Its notable limitation: no speaker diarization. There is no mention of it anywhere in the repository, despite many articles claiming otherwise.

faster-whisper: the speed choice

A reimplementation using CTranslate2. 25,187 stars, MIT. Requires Python 3.9+.

This is what to use if you have a GPU and want maximum throughput. The int8 configuration is the standout: 2.4x faster than the original using 40% less memory, which also means larger models fit on smaller cards.

No diarization built in, though it integrates with WhisperX for that.

whisper.cpp: the runs-anywhere choice

A port to plain C/C++ with no dependencies. 53,365 stars, MIT.

Its reach is the point: macOS on Intel and Apple Silicon, iOS, Android, Linux, FreeBSD, WebAssembly, Windows, Docker, and Raspberry Pi. It uses integer quantisation and the ggml format to run on hardware nothing else will touch.

On Apple Silicon, building with Core ML support runs the encoder on the Neural Engine for more than 3x faster performance than CPU alone.

It includes tinydiarize for speaker separation, though that remains experimental.

Use it for CPU-only machines, embedded hardware, or anywhere you want a single binary with no Python environment.

Buzz: the one for humans

MIT licensed, and the answer for anyone who does not want a terminal.

Buzz is a genuine desktop application. Installers for Apple Silicon Macs and Windows, drag-and-drop transcription, live capture from your microphone, export to TXT, SRT and VTT, and speaker identification built in. It uses whisper.cpp with Vulkan, CUDA, Apple Silicon and Hugging Face Transformers as backends.

Two honest friction points: Intel Mac support ended at v1.4.5, and the Windows installer is unsigned, so you will click through a SmartScreen warning.

For most people reading this, Buzz is the right starting point.

WhisperX: the diarization choice

BSD-2-Clause licensed. Built specifically to add speaker diarization using pyannote's community model, which is CC-BY-4.0 and permits commercial use with attribution.

It requires a Hugging Face token and accepting the model's user agreement. It claims 70x realtime with large-v2 under 8GB of VRAM.

Use it when knowing who said what actually matters.

Which model size

Independent of implementation, Whisper comes in sizes:

ModelParamsVRAMRelative speed
tiny39M~1GB~10x
base74M~1GB~7x
small244M~2GB~4x
medium769M~5GB~2x
large1550M~10GB1x
turbo809M~6GB~8x

Turbo is the practical default: an optimised large-v3 giving most of the accuracy at roughly eight times the speed.

One caveat: turbo is not trained for translation. It transcribes in the source language but will not translate to English. Use large for that.

English-only variants (tiny.en, base.en, small.en) outperform the multilingual versions at smaller sizes, though the gap narrows at medium and above.

Combining choices

Implementation and model size multiply. faster-whisper at int8 running the turbo model is dramatically quicker than openai/whisper running large, and the quality difference is small.

A reasonable default for most hardware: faster-whisper with turbo at int8 if you have a GPU, or whisper.cpp with small or medium if you do not.

Which should you choose?

Choose Buzz if you want a GUI, installers and speaker identification with no setup. Right for most people.

Choose faster-whisper if you have a GPU, are comfortable with Python, and want maximum speed.

Choose whisper.cpp if you are CPU-only, on unusual hardware, or want a dependency-free binary.

Choose WhisperX if speaker diarization is the requirement.

Choose openai/whisper if you want the reference implementation for experimentation.

For the full picture including paid alternatives, see our guide to free AI transcription tools.

The verdict

Whisper's reputation for being slow comes almost entirely from people using the original Python implementation on modest hardware.

Switching to faster-whisper at int8 more than halves the time and cuts memory by 40% for identical output. Switching to Buzz removes the terminal entirely. Neither costs anything, and both are MIT licensed. The tooling around this model got very good, and most guides have not caught up.

Related reading: free AI transcription tools and Whisper versus Otter and Fireflies.

Frequently Asked Questions

What is the fastest Whisper implementation?

faster-whisper at int8 precision. Benchmarked on 13 minutes of audio with large-v2 on an RTX 3070 Ti, it took 59 seconds using 2,926MB, against openai/whisper at 2m23s and 4,708MB. That is 2.4x faster on 40% less memory with identical output.

Which Whisper version works without a GPU?

whisper.cpp, which is written in plain C/C++ with no dependencies. It runs on macOS, Windows, Linux, iOS, Android, WebAssembly and even Raspberry Pi. On Apple Silicon, building with Core ML runs the encoder on the Neural Engine for more than 3x faster performance than CPU alone.

What is the easiest way to use Whisper?

Buzz, an MIT-licensed desktop app with installers for Apple Silicon Macs and Windows. It offers drag-and-drop transcription, live microphone capture, speaker identification and export to TXT, SRT and VTT with no terminal required.

Which Whisper model size should you use?

Turbo for most people. It is an optimised large-v3 giving most of the accuracy at roughly eight times the speed, needing about 6GB of VRAM. Note turbo is not trained for translation, so use large if you need to translate to English.

Which Whisper tools support speaker diarization?

WhisperX gives the best quality using pyannote, under a BSD-2-Clause licence. Buzz includes speaker identification with a GUI. whisper.cpp has experimental tinydiarize. The original openai/whisper does not support diarization at all.