This project provides pre-compiled, ready-to-use binaries for ggerganov/whisper.cpp. It aims to help users quickly and easily deploy and use the Whisper ASR model on various mainstream operating systems without the need to configure a complex build environment.
This project is for you if you are:
- Developers who want to quickly integrate Whisper's speech recognition capabilities into their applications or services without spending time on C++ environment setup and compilation.
- General users who want to run a local, offline speech-to-text service on their personal computer to protect data privacy.
- Tech enthusiasts interested in AI models who want to experience the power of Whisper on their own devices.
whisper.cpp is a high-performance C++ port of OpenAI's Whisper model. Whisper is a powerful general-purpose Automatic Speech Recognition (ASR) model that can accurately transcribe audio from various languages into text.
The core advantages of the whisper.cpp project are:
- Pure C/C++ Implementation: No external dependencies, making it easy to integrate and deploy.
- Highly Optimized: Supports various instruction sets like Apple Silicon (Metal), AVX, AVX2, and AVX512 for optimal performance.
- Cross-Platform: Runs on various platforms including Windows, macOS, Linux, Android, and iOS.
- Low Resource Consumption: Optimized for memory and processor usage, allowing it to run efficiently on resource-constrained devices.
The Whisper model comes in various sizes, such as tiny, base, small, medium, and large. The whisper.cpp project excels at running these models efficiently on CPUs.
Small models (like tiny and small) are ideal for on-device deployment. Their advantages include:
- Extremely Low Resource Usage: They require very little memory and CPU, so they won't burden your device.
- Fast Response Speed: Suitable for real-time transcription scenarios that require immediate feedback.
- Completely Offline: All computations are done locally, ensuring data privacy and security.
For most personal applications and edge integration scenarios, small models strike an excellent balance between performance, speed, and resource consumption.
This project automatically builds and verifies binaries for the following platforms via CI/CD:
| Operating System | Architecture | Notes |
|---|---|---|
| Linux | x86_64 (amd64) | Built on the latest Ubuntu |
| Windows | x86_64 (amd64) | Built on the latest Windows |
| macOS | x86_64 (Intel) | Built on the latest macOS |
| macOS | ARM64 (Apple Silicon) | Built with Metal support for optimal GPU acceleration |
Go to the download page and get the .zip package corresponding to your OS and architecture.
Download URL: https://yaklang.oss-accelerate.aliyuncs.com/whisper.cpp/
Please navigate to the folder with the most recent date (e.g., 20231027/) to download. We also provide .sha256 files for you to verify file integrity.
You will need a model file in ggml format. You can download one from Hugging Face.
Here is an example of downloading the tiny model:
# Create a directory for models
mkdir models
# Download the tiny-q5 model
curl -L -o models/ggml-tiny.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.binUnzip the downloaded package and use the following command to start the whisper-server:
# Make the binary executable (Linux / macOS)
chmod +x ./whisper-server
# Start the server, specifying the model path with the -m flag
./whisper-server -m ./models/ggml-tiny.binThe server listens on port 8080 by default.
You can use curl or any HTTP client to send a multipart/form-data request to the /inference endpoint for speech recognition.
curl --request POST \
--url http://127.0.0.1:8080/inference \
--header 'Content-Type: multipart/form-data' \
--form 'file=@/path/to/your/audio.mp3'The server will return a JSON result containing the transcribed text.
This project only provides pre-compiled binaries for whisper.cpp to facilitate user access. All core code and functionality come from the official ggerganov/whisper.cpp repository. For source code, feature requests, or to report core issues, please visit the official project.
本项目为 ggerganov/whisper.cpp 项目提供预编译的、可直接执行的二进制文件，旨在帮助用户快速、轻松地在各种主流操作系统上部署和使用 Whisper ASR 模型，无需自行配置复杂的编译环境。
如果您是以下用户之一，本项目将为您带来便利：
- 开发者：希望将 Whisper 的语音识别能力快速集成到自己的应用或服务中，而不想花费时间在 C++ 环境的搭建和编译上。
- 普通用户：希望在个人电脑上运行一个本地的、离线的语音转文字服务，以保护数据隐私。
- 技术爱好者：对 AI 模型感兴趣，希望在自己的设备上体验 Whisper 的强大功能。
whisper.cpp 是 OpenAI Whisper 模型的高性能 C++ 移植版本。Whisper 是一个强大的通用自动语音识别（ASR）模型，可以准确地将各种语言的音频转录为文字。
whisper.cpp 项目的核心优势在于：
- 纯 C/C++ 实现：无任何外部依赖，易于集成和部署。
- 极致优化：支持 Apple Silicon (Metal)、AVX、AVX2、AVX512 等多种指令集，以实现最佳性能。
- 跨平台：可在 Windows、macOS、Linux、Android、iOS 等多种平台上运行。
- 低资源消耗：对内存和处理器占用进行了优化，使其可以在资源受限的设备上高效运行。
Whisper 模型提供多种尺寸，如 tiny、base、small、medium、large。whisper.cpp 项目尤其擅长在 CPU 上高效运行这些模型。
小模型（如 tiny 和 small）是本地设备部署的理想选择。它们的优势在于：
- 极低的资源占用：运行时仅需很少的内存和 CPU，不会对您的设备造成负担。
- 快速的响应速度：适合需要即时反馈的实时转录场景。
- 完全离线：所有计算都在本地完成，确保了数据隐私和安全。
对于大多数个人应用和端侧集成场景，小模型在性能、速度和资源消耗之间取得了绝佳的平衡。
本项目通过 CI/CD 自动构建并验证了以下平台的二进制文件：
| 操作系统 | 架构 | 说明 |
|---|---|---|
| Linux | x86_64 (amd64) | 在 Ubuntu 最新版上构建 |
| Windows | x86_64 (amd64) | 在 Windows 最新版上构建 |
| macOS | x86_64 (Intel) | 在 macOS 最新版上构建 |
| macOS | ARM64 (Apple Silicon) | 启用 Metal 支持以获得最佳 GPU 加速性能 |
前往下载页面，根据您的操作系统和架构下载对应的 .zip 压缩包。
下载地址: https://yaklang.oss-accelerate.aliyuncs.com/whisper.cpp/
请进入最新日期的文件夹（例如 20231027/）进行下载。我们同时提供了 .sha256 文件供您校验文件完整性。
您需要一个 ggml 格式的模型文件。您可以从 Hugging Face 上下载。
以下是一个 tiny 模型的下载示例：
# 创建一个用于存放模型的文件夹
mkdir models
# 下载 tiny-q5 模型
curl -L -o models/ggml-tiny.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin解压下载的压缩包，并使用以下命令启动 whisper-server 服务：
# 赋予可执行权限 (Linux / macOS)
chmod +x ./whisper-server
# 启动服务，-m 参数指定模型路径
./whisper-server -m ./models/ggml-tiny.bin服务默认在 8080 端口上监听。
您可以使用 curl 或任何 HTTP 客户端向 /inference 端点发送一个 multipart/form-data 请求来进行语音识别。
curl --request POST \
--url http://127.0.0.1:8080/inference \
--header 'Content-Type: multipart/form-data' \
--form 'file=@/path/to/your/audio.mp3'服务器将返回一个包含识别文本的 JSON 结果。
本项目仅提供 whisper.cpp 的预编译二进制文件，以方便用户使用。所有核心代码和功能均来自 ggerganov/whisper.cpp 官方仓库。如需查看源码、提出功能建议或报告核心问题，请访问官方项目。