transcribe.cpp
SHA-256A C/C++ speech-to-text inference library supporting 16+ model families via GGUF/ggml with GPU acceleration (Metal, Vulkan, CUDA) and a fast CPU path.
Smart Download
Download Download Version
v0.1.3 · 24.9 MB
A lightweight, high-performance speech-to-text engine supporting 16+ model families with GPU acceleration.
Core Features
- Supports 16+ model families (Parakeet, Whisper, Canary, Moonshine, etc.) with 60+ variants
- GPU acceleration via Metal, Vulkan, CUDA; CPU fallback with tinyBLAS
- Official bindings for Python, TypeScript, Rust, and Swift
- Numerical verification and WER testing against reference implementations
- Streaming and batch inference support
What It Can't Do
- •Input audio must be 16 kHz mono WAV; use ffmpeg/sox to convert other formats
- •CPU inference benefits greatly from libopenblas-dev (10-15x decoder speedup)
- •Quantized models (e.g., Q4_K_M) may slightly reduce accuracy; prefer F16 for critical tasks
- •Speaker diarization is not supported except in the MOSS Transcribe-Diarize variant
Use Cases
- Embedded offline speech-to-text in desktop or mobile applications
- High-throughput server-side transcription with GPU acceleration
Detailed Introduction
transcribe.cpp is a lightweight, high-performance speech-to-text inference engine written in C/C++. It runs over 60 variants from 16 model families—including Parakeet, Whisper, Canary, Moonshine, SenseVoice, and more—using the GGUF format on top of the ggml runtime. Unlike monolithic solutions like OpenAI Whisper (which is Python-heavy and requires large dependencies), transcribe.cpp is a tiny, embeddable library with official bindings for Python, TypeScript, Rust, and Swift. It leverages Metal (Apple Silicon), Vulkan, and CUDA for GPU inference, and includes a tinyBLAS-accelerated CPU fallback. Every published model is numerically verified and WER-tested against reference implementations, ensuring production-grade accuracy. The project is MIT-licensed and backed by Mozilla AI, Hugging Face, and Modal.
Tags
Getting Started
Download installer
Click the button above to download the installer for your system
Install the software
Install the appropriate package for your distro (dpkg / rpm / AppImage)
Download pre-built binaries or build from source (cmake -B build && cmake --build build)
Download GGUF model weights from Hugging Face (handy-computer org)
Run transcribe-cli -m model.gguf input.wav (16kHz mono WAV required)
- Download pre-built binaries or build from source (cmake -B build && cmake --build build)
- Download GGUF model weights from Hugging Face (handy-computer org)
- Run transcribe-cli -m model.gguf input.wav (16kHz mono WAV required)
SHA-256 checksum verified
Checksum extracted from GitHub official Release page
SHA256 Checksum
94c3914b2f8648a1d7c61a713b36ab05dfc38b8e6781ae13ecf841b559c77bffThis checksum is extracted from the GitHub Release page. Verify file integrity after download.
All SHA-256 checksums on this platform are extracted from the project's official GitHub Release page, without any modification. You can independently verify them on the GitHub Releases page.
Open Source Transparency
View GitHub SourceUninstall Info
Delete the project directory and build artifacts. If installed via package manager, use the corresponding uninstall command.
No Extra Dependencies
Ready to use after download. No additional runtime required.
Similar Projects
ollama
Ollama lets you download, run, and manage large language models locally. One command, multiple platforms, endless possibilities.
llama.cpp
High-performance LLM inference engine in C/C++ with minimal dependencies, supporting quantized models (1.5–8 bit) and diverse hardware (Apple Silicon, CUDA, Vulkan, etc.).
opencv
OpenCV is an open-source computer vision and machine learning library with over 2500 optimized algorithms for real-time image and video analysis.