faster-gigaam

Optimized inference engine for GigaAM ASR models with CUDA acceleration, batch processing, and unlimited audio length support.

Features

• 🚀 CUDA Acceleration: 5-13x faster than CPU with mixed precision inference
• 📦 Batch Processing: Process multiple audio chunks simultaneously for maximum GPU utilization
• ♾️ Unlimited Audio Length: Automatic chunking with overlap - no 25-second limit
• 🎯 Simple API: Single transcribe() method handles any audio length
• ⚡ High Throughput: Process up to 262 seconds of audio per second (262x realtime on GPU)
• 🔧 Zero Extra Dependencies: No VAD or additional models required

Performance

Benchmark results on real audio (v3_ctc model):

Test Hardware:

• CPU: Intel Core i9-9900K
• RAM: 32GB
• GPU: NVIDIA RTX 3080 (10GB VRAM)
• CUDA: 12.8

Short Audio (11.29s)

Implementation	Device	Batch Size	Time (s)	Speedup vs Original CPU
Original GigaAM	CPU	N/A	0.537	1.00x (baseline)
Original GigaAM	CUDA	N/A	0.122	4.40x
faster-gigaam	CPU	8	0.513	1.05x
faster-gigaam	CUDA	8	0.111	4.82x

Long Audio (70.51s)

Implementation	Device	Batch Size	Time (s)	Throughput
Original GigaAM	CPU	N/A	❌ Failed	Requires transcribe_longform
Original GigaAM	CUDA	N/A	❌ Failed	Requires transcribe_longform
faster-gigaam	CPU	8	3.248	21.7x realtime
faster-gigaam	CUDA	8	0.269	262x realtime

Key Advantage: faster-gigaam handles long audio out-of-the-box, while original GigaAM requires additional dependencies (pyannote.audio, HuggingFace token) and a separate transcribe_longform method.

Installation

Note: PyTorch is not included as a dependency to allow you to choose between CPU and CUDA versions. Install PyTorch first according to your hardware.

Tip: For faster installations, consider using UV package manager instead of pip.

Step 1: Install PyTorch

Choose the appropriate PyTorch version for your system:

With CUDA (for GPU acceleration):

# CUDA 12.8 (recommended)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

# Or with UV (faster package manager)
uv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

CPU only:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

See PyTorch installation guide for more options.

Step 2: Install GigaAM

pip install git+https://github.com/salute-developers/GigaAM.git

Step 3: Install faster-gigaam

# Clone the repository
git clone https://github.com/yourusername/faster-gigaam.git
cd faster-gigaam

# Install in development mode
pip install -e .

Requirements:

• Python >= 3.10 (tested with Python 3.13)
• PyTorch >= 2.0.0 (installed in Step 1)
• GigaAM (installed in Step 2)

Recommended:

• UV package manager for faster installations

Quick Start

from faster_gigaam import FastGigaAM

# Initialize model with GPU and batch processing
model = FastGigaAM(
    model_name="v3_ctc",
    device="cuda",
    batch_size=8
)

# Transcribe audio (any length!)
segments, info = model.transcribe("long_audio.wav")

# Print results
for segment in segments:
    print(f"[{segment.start:.2f}s - {segment.end:.2f}s] {segment.text}")

# Print performance info
print(f"\nProcessed {info.duration:.1f}s audio in {info.processing_time:.2f}s")
print(f"Throughput: {info.duration / info.processing_time:.1f}x realtime")

CPU Usage

# Use CPU if CUDA is not available
model = FastGigaAM(
    model_name="v3_ctc",
    device="cpu",
    batch_size=4
)

segments, info = model.transcribe("audio.wav")

Batch Processing Multiple Files

# Process multiple files efficiently
audio_files = ["file1.wav", "file2.wav", "file3.wav"]
results = model.transcribe_batch(audio_files)

for i, (segments, info) in enumerate(results):
    print(f"\nFile {i+1}:")
    for segment in segments:
        print(f"  [{segment.start:.2f}s - {segment.end:.2f}s] {segment.text}")

Documentation

• Examples: See examples/ folder for more usage examples
• Benchmarks: See BENCHMARK_RESULTS.md for detailed performance analysis
• API Reference: See docstrings in faster_gigaam/fast_gigaam.py

Supported Models

All GigaAM ASR models are supported:

• v3_ctc - CTC-based model
• v3_e2e_rnnt - RNN-T based model
• Other GigaAM model variants

Project Status

This is a pet project created for personal use and shared with the community.

• ✅ Fully functional and tested
• ⚠️ Updates will be sporadic based on my needs
• 🤝 Issues and PRs welcome, but no guarantees on response time
• 🍴 Feel free to fork if you need active maintenance

License

MIT License - same as GigaAM

Acknowledgments

Built on top of the excellent GigaAM project by Salute Developers.