faster-whisper

What it is and what it does

Faster-whisper is a reimplementation of OpenAI's Whisper speech-to-text model using CTranslate2, a fast inference engine for Transformer models. It trades the original Whisper library for a more efficient backend, achieving measurable speed and memory improvements on both CPU and GPU while maintaining transcription accuracy. The package handles audio decoding via PyAV (bundled FFmpeg), so you don't need to install FFmpeg separately.

The package supports multiple precision modes (fp32, fp16, int8) and batch processing, allowing you to tune speed and memory trade-offs for your hardware. It can run on CPU or GPU, and works with Whisper's standard model sizes (tiny, base, small, medium, large) as well as Distil-Whisper checkpoints. The transcription API returns segments with timestamps and detected language information.

Use it for:

• Transcribe long audio or video files on GPU with reduced latency and VRAM compared to openai/whisper.
• Run speech-to-text on CPU with int8 quantization to fit within constrained memory budgets.
• Batch-process multiple audio files in parallel using the BatchedInferencePipeline for throughput.
• Deploy Whisper transcription in production where inference speed and memory efficiency are critical.
• Use Distil-Whisper checkpoints for faster, lighter-weight transcription with acceptable accuracy trade-off.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Transcribes audio to text using OpenAI's Whisper model, reimplemented with CTranslate2 for faster inference and lower memory use than the original.

Yes, if you need Whisper transcription and speed or memory efficiency matters. The package is mature (Beta status, 24910 stars, no known vulnerabilities), permissively licensed, and has low install friction. Maintenance is aging (287 days since last release), but the repository remains active. GPU users must have CUDA 12 libraries available; CPU-only use is simpler. Start here if you're choosing between openai/whisper and faster-whisper for the same accuracy at lower cost.

Install

faster-whisper on PyPI

pip install faster-whisper

uv add faster-whisper

poetry add faster-whisper

Installing faster-whisper

Before you install

Low friction install with a pure-Python wheel. Maintenance is aging—last release was 287 days ago—but the repository remains active with recent commits and substantial community engagement (24910 stars). Six runtime dependencies are all established packages.

License in practice

MIT license is permissive; you can use this package freely in commercial and private projects with minimal restrictions.

Quickstart

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("large-v3", device="cuda", compute_type="float16")
segments, info = model.transcribe("audio.mp3", beam_size=5)
for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

GPU execution requires NVIDIA cuBLAS and cuDNN libraries for CUDA 12; CPU-only use works without them. Python 3.9 or greater required.

Verify before relying

• Whether the 4x speedup claim applies to all model sizes and hardware configurations, or only specific benchmarked scenarios.
• Real-world accuracy parity with openai/whisper across diverse audio types and languages.
• Whether batched inference (batch_size parameter) is stable and recommended for production use.

Package facts

Evidence: faster_whisper-1.2.1-py3-none-any.whl

Keywords: openai, whisper, speech, ctranslate2, inference, quantization, transformer

Tags

More Artificial Intelligence packages

Further reading