whisper.py 0.1.0

A Python wrapper for whisper.cpp - fast automatic speech recognition

Whispy - Fast Speech Recognition CLI

A fast and efficient command-line interface for whisper.cpp, providing automatic speech recognition with GPU acceleration.

Features

• 🚀 Fast transcription using whisper.cpp with GPU acceleration (Metal on macOS, CUDA on Linux/Windows)
• 🎯 Simple CLI interface for easy audio transcription
• 📁 Multiple audio formats supported (WAV, MP3, FLAC, OGG)
• 🌍 Multi-language support with automatic language detection
• 📝 Flexible output options (stdout, file)
• 🔧 Auto-detection of models and whisper-cli binary
• 🏗️ Automatic building of whisper.cpp if needed

Installation

Quick Install (Recommended)

Install directly from GitHub with automatic setup:

pip install git+https://github.com/amarder/whispy.git

This will automatically:

• Clone whisper.cpp to ~/.whispy/whisper.cpp
• Build the whisper-cli binary with GPU acceleration
• Install the whispy CLI

Manual Install

If you prefer to install manually:

Prerequisites

• Python 3.7+
• CMake 3.10+ (for building whisper.cpp)
• C++ compiler with C++17 support
• Git (for cloning whisper.cpp)

Steps

# Clone repository
git clone https://github.com/amarder/whispy.git
cd whispy

# Install whispy
pip install -e .

# Clone whisper.cpp if you don't have it
git clone https://github.com/ggerganov/whisper.cpp.git

# Build whisper-cli (or use: whispy build)
cd whisper.cpp
cmake -B build
cmake --build build -j --config Release
cd ..

Requirements

Basic requirements:

• Python 3.7+
• CMake (for building whisper.cpp)
• C++ compiler (gcc, clang, or MSVC)
• Git

For audio recording features:

• Microphone access
• Audio drivers (pre-installed on most systems)
• Additional Python packages: sounddevice, numpy, scipy

Supported platforms:

• 🍎 macOS (Intel & Apple Silicon) with CoreAudio
• 🐧 Linux (with ALSA/PulseAudio)
• 🪟 Windows (with DirectSound)

Download a model

After installation, download a model to use for transcription:

# For pip installs from GitHub
cd ~/.whispy/whisper.cpp
sh ./models/download-ggml-model.sh base.en

# For manual installs
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en

# Alternative: Download directly to models/
mkdir -p models
curl -L -o models/ggml-base.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

Usage

Basic transcription

# Transcribe an audio file
whispy transcribe audio.wav

# Transcribe with explicit model
whispy transcribe audio.wav --model models/ggml-base.en.bin

# Transcribe with language specification
whispy transcribe audio.wav --language en

# Save transcript to file
whispy transcribe audio.wav --output transcript.txt

# Verbose output
whispy transcribe audio.wav --verbose

Record and transcribe

Record audio from your microphone and transcribe it in real-time:

# Record and transcribe (press Ctrl+C to stop recording)
whispy record-and-transcribe

# Test microphone before recording
whispy record-and-transcribe --test-mic

# Record with specific model and language
whispy record-and-transcribe --model models/ggml-base.en.bin --language en

# Save both transcript and audio
whispy record-and-transcribe --output transcript.txt --save-audio recording.wav

# Verbose output with device information
whispy record-and-transcribe --verbose

Real-time transcription

Transcribe audio from your microphone in real-time using streaming chunks:

# Start real-time transcription (press Ctrl+C to stop)
whispy realtime

# With custom settings for faster/slower processing
whispy realtime --chunk-duration 2.0 --overlap-duration 0.5 --silence-threshold 0.02

# Show individual chunks instead of continuous output
whispy realtime --show-chunks

# Save final transcript to file
whispy realtime --output live_transcript.txt

# Test real-time setup
whispy realtime --test-setup

# Verbose mode for debugging
whispy realtime --verbose

Real-time Parameters:

• --chunk-duration: Duration of each audio chunk in seconds (default: 3.0)
• --overlap-duration: Overlap between chunks in seconds (default: 1.0)
• --silence-threshold: Voice activity detection threshold (default: 0.01)
• --show-chunks: Show individual chunk transcripts instead of continuous mode
• --test-setup: Test real-time setup without starting transcription

System information

# Check system status
whispy info

# Show version
whispy version

# Build whisper-cli if needed
whispy build

Supported audio formats

• WAV
• MP3
• FLAC
• OGG

Available models

Download models using whisper.cpp's script or directly:

• tiny.en, tiny - Fastest, least accurate
• base.en, base - Good balance of speed and accuracy
• small.en, small - Better accuracy
• medium.en, medium - High accuracy
• large-v1, large-v2, large-v3 - Best accuracy, slower

Examples

# Quick transcription with auto-detected model
whispy transcribe meeting.wav

# High-quality transcription
whispy transcribe interview.mp3 --model whisper.cpp/models/ggml-large-v3.bin

# Transcribe non-English audio
whispy transcribe spanish_audio.wav --language es

# Save results and show details
whispy transcribe podcast.mp3 --output transcript.txt --verbose

# Record and transcribe in real-time
whispy record-and-transcribe

# Record with high-quality model and save everything
whispy record-and-transcribe \
  --model whisper.cpp/models/ggml-large-v3.bin \
  --output meeting-notes.txt \
  --save-audio meeting-recording.wav \
  --verbose

# Quick voice memo transcription
whispy record-and-transcribe --language en --output memo.txt

# Real-time transcription with live output
whispy realtime

# Real-time transcription with custom settings
whispy realtime --chunk-duration 2.0 --show-chunks --output live_notes.txt

Testing

Whispy includes a comprehensive test suite to ensure the CLI works correctly with different scenarios.

Running Tests

# Install development dependencies
pip install -e ".[dev]"

# Run all tests
pytest

# Run tests with verbose output
pytest -v

# Run only unit tests
pytest tests/test_unit.py

# Run only CLI tests
pytest tests/test_cli.py

# Run tests with coverage
pytest --cov=whispy --cov-report=html

# Skip slow tests
pytest --fast

Test Categories

• Unit tests (tests/test_unit.py): Test individual functions and modules
• CLI tests (tests/test_cli.py): Test command-line interface functionality
• Integration tests: Test full workflows with real audio files

Using the Test Runner

# Use the convenience script
python run_tests.py --help

# Run unit tests only
python run_tests.py -t unit -v

# Run with coverage
python run_tests.py -c -v

# Run fast tests only
python run_tests.py -f

Test Requirements

• pytest >= 7.0.0
• pytest-cov >= 4.0.0
• pytest-mock >= 3.10.0
• Sample audio files (JFK sample from whisper.cpp)

What's Tested

• ✅ CLI commands (help, version, info, transcribe, record-and-transcribe)
• ✅ Audio file transcription with sample files
• ✅ Audio recording from microphone
• ✅ Real-time record-and-transcribe workflow
• ✅ Microphone testing functionality
• ✅ Error handling for invalid files/models/devices
• ✅ Output file generation
• ✅ Language options and verbose modes
• ✅ System requirements and binary detection
• ✅ Model file discovery and validation

Development

Project Structure

whispy/
├── whispy/
│   ├── __init__.py       # Package initialization
│   ├── cli.py           # Command-line interface
│   └── transcribe.py    # Core transcription logic
├── whisper.cpp/         # Git submodule (whisper.cpp source)
├── models/              # Model files directory
├── pyproject.toml       # Project configuration
└── README.md

How it works

Whispy works as a wrapper around the whisper-cli binary from whisper.cpp:

1. Auto-detection: Finds whisper-cli binary and model files automatically
2. Subprocess calls: Runs whisper-cli as a subprocess for transcription
3. Output parsing: Captures and returns the transcribed text
4. Performance: Gets full GPU acceleration and optimizations from whisper.cpp

Building from source

# Clone with whisper.cpp submodule
git clone --recursive https://github.com/your-username/whispy.git
cd whispy

# Install in development mode
pip install -e .

# Build whisper.cpp
whispy build
# OR manually:
# cd whisper.cpp && cmake -B build && cmake --build build -j --config Release

Adding new features

The CLI is built with Typer and can be easily extended:

@app.command()
def new_command():
    """Add a new command to the CLI"""
    console.print("New feature!")

Performance

Whispy automatically uses the best available backend:

• macOS: Metal GPU acceleration
• Linux/Windows: CUDA GPU acceleration (if available)
• Fallback: Optimized CPU with BLAS

Typical performance on Apple M1:

• ~10x faster than real-time for base.en model
• ~5x faster than real-time for large-v3 model

Troubleshooting

whisper-cli not found

# Check if whisper-cli exists
whispy info

# Build whisper-cli
whispy build

# Or build manually
cd whisper.cpp
cmake -B build && cmake --build build -j --config Release

No model found

# Download a model
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en

# Or specify model explicitly
whispy transcribe audio.wav --model /path/to/model.bin

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Development setup

git clone --recursive https://github.com/your-username/whispy.git
cd whispy
pip install -e .
whispy build

Acknowledgments

• whisper.cpp - Fast C++ implementation of OpenAI's Whisper
• OpenAI Whisper - Original Whisper model
• Typer - CLI framework