TL;DR — Quick Summary

Run OpenAI's Whisper speech-to-text locally for private, free transcription. Covers CLI, Docker, GPU, Whisper.cpp for CPU, faster-whisper, and web UI.

What Is Whisper?

Whisper is OpenAI’s open-source automatic speech recognition (ASR) model. It can transcribe audio in 99 languages, translate speech to English, and generate subtitles with accurate timestamps — all running locally on your own hardware, completely free and private.

Key features:

  • 99 languages — transcripts in the original language or translated to English
  • Multiple model sizes — from tiny (75MB) to large-v3 (3GB)
  • Subtitle generation — SRT, VTT, and TSV formats with timestamps
  • No API needed — runs entirely offline
  • GPU acceleration — CUDA for NVIDIA GPUs
  • Multiple implementations — Python, C++ (whisper.cpp), faster-whisper

Model Comparison

Model Size VRAM English WER Speed (GPU) Speed (CPU)
tiny 75 MB ~1 GB 8.0% ~32x realtime ~10x realtime
base 142 MB ~1 GB 5.7% ~16x realtime ~7x realtime
small 466 MB ~2 GB 4.2% ~6x realtime ~2x realtime
medium 1.5 GB ~5 GB 3.5% ~2x realtime ~0.5x realtime
large-v3 3 GB ~6 GB 2.9% ~1x realtime ~0.1x realtime

Tip: For most use cases, small or medium offers the best accuracy vs speed tradeoff. Use large-v3 only when accuracy is critical.

Installation

Python (Original)

pip install openai-whisper

# Transcribe
whisper audio.mp3 --model base

# Transcribe with language detection and SRT output
whisper interview.wav --model small --output_format srt

# Translate to English
whisper audio_spanish.mp3 --model medium --task translate

faster-whisper (4x Faster)

pip install faster-whisper

python -c "
from faster_whisper import WhisperModel
model = WhisperModel('base', 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}')
"

whisper.cpp (Best for CPU)

git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp && make

# Download a model
bash models/download-ggml-model.sh base

# Transcribe (WAV format required)
./main -m models/ggml-base.bin -f audio.wav -otxt -osrt

Docker

docker run --gpus all -v $(pwd):/data \
  onerahmet/openai-whisper-asr-webservice:latest \
  whisper /data/audio.mp3 --model base --output_dir /data/

Web UI Options

For a browser-based interface:

Project Description Docker
Whisper Web Simple upload and transcribe UI docker run -p 9000:9000 pluja/whishper
Whisper ASR Webservice REST API with Swagger UI onerahmet/openai-whisper-asr-webservice
Subtitle Edit Full editor with Whisper integration Desktop app

Common Use Cases

Use Case Model Command
Meeting transcription small whisper meeting.mp3 --model small
Video subtitles medium whisper video.mp4 --model medium --output_format srt
Podcast transcription base whisper podcast.mp3 --model base --output_format txt
Translate foreign audio medium whisper foreign.mp3 --model medium --task translate
Batch process folder base for f in *.mp3; do whisper "$f" --model base; done

Whisper vs. Cloud Speech-to-Text

Aspect Whisper (Local) Google STT AWS Transcribe
Cost Free $0.006-0.048/min $0.024/min
Privacy ✅ Data stays local ❌ Cloud ❌ Cloud
Offline ✅ ❌ ❌
Languages 99 125+ 100+
Accuracy Excellent Excellent Good
Custom vocabulary ❌ ✅ ✅
Real-time streaming Limited ✅ ✅
Speaker diarization Via plugin ✅ ✅

Summary

Whisper gives you state-of-the-art speech-to-text transcription running locally, privately, and for free. Use the Python package for quick transcriptions, faster-whisper for GPU-accelerated performance, or whisper.cpp for efficient CPU-only operation.