FunASR 跑进 llama.cpp:中文语音识别的 whisper.cpp 替代品(CPU、零 Python、下载即用)
- + -whisper.cpp 是端侧语音识别的事实标准——单个自包含二进制、CPU 就能跑、零依赖。但 Whisper 在中文上偏弱。现在 FunASR 有了 llama.cpp / GGUF 运行时:一样的"下载即用"体验,**单个静态二进制、零 Python、内置 VAD、吃任意音频**,而且在中文 CPU 上比 whisper.cpp 准约 2.7 倍。
+whisper.cpp 是端侧语音识别的事实标准——单个自包含二进制、CPU 就能跑、零依赖。但 Whisper 在中文上偏弱。现在 FunASR 有了 llama.cpp / GGUF 运行时:一样的"下载即用"体验,单个静态二进制、零 Python、内置 VAD、吃任意音频,而且在中文 CPU 上比 whisper.cpp 准约 2.7 倍。
-👉 想直接下载? 各平台预编译二进制(Linux/macOS/Windows)+ 一站式 quickstart 与 benchmark 见 在 CPU 上运行 FunASR(llama.cpp 下载页)。
+👉 想直接下载? v0.2.0 提供 9 个 Linux、macOS、Windows 预编译包及 SHA-256;下载矩阵、quickstart 与已知限制见 llama.cpp / GGUF 部署页。
3 步跑通(全部实测)
# 1. 下预编译二进制(linux-x64 / linux-arm64 / macos-arm64 / windows-x64) -wget https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.1.9/funasr-llamacpp-linux-x64-vulkan.tar.gz +wget https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-x64-vulkan.tar.gz tar xzf funasr-llamacpp-linux-x64-vulkan.tar.gz # Windows Vulkan 包: funasr-llamacpp-windows-x64-vulkan.zip(运行时加 --backend vulkan) @@ -90,7 +98,8 @@你能得到什么
- 内置 FSMN-VAD:长音频自动分段(`--vad fsmn-vad.gguf`),裸二进制即达参考精度,无需 Python 前端
- 任意音频:wav / mp3 / flac、任意采样率/声道,二进制内部重采样 -
- 多平台预编译:Linux x64/arm64、macOS arm64、Windows x64,含 Linux/Windows Vulkan 与 Windows CUDA 加速包(Release);其它平台一条 CMake 命令自构建 +
- 多平台预编译:v0.2.0 提供 9 个 Linux x64/arm64、macOS arm64、Windows x64 包,含 Linux/Windows Vulkan 与 Windows CUDA 加速包(Release);其它平台一条 CMake 命令自构建 +
- 直接输出字幕:Fun-ASR-Nano、SenseVoiceSmall 和 Paraformer CLI 均可通过
--srt输出带时间戳的 SRT 字幕
中文 CPU 上 vs whisper.cpp
diff --git a/web-pages/product-site/legacy/en/blog/funasr-llama-cpp-whisper-cpp-alternative.html b/web-pages/product-site/legacy/en/blog/funasr-llama-cpp-whisper-cpp-alternative.html index e52113009..3c174ad89 100644 --- a/web-pages/product-site/legacy/en/blog/funasr-llama-cpp-whisper-cpp-alternative.html +++ b/web-pages/product-site/legacy/en/blog/funasr-llama-cpp-whisper-cpp-alternative.html @@ -18,7 +18,7 @@ +footer a{color:#94a3b8} +@media(max-width:760px){ +.nav{padding:12px 0} +.nav .container{gap:12px;flex-wrap:nowrap} +.nav-links{display:none} +.nav-logo{margin-right:auto} +.nav-btn{padding:7px 12px} +article{padding-top:96px} +}FunASR on llama.cpp — a whisper.cpp Alternative for Chinese ASR (CPU, no Python)
- +whisper.cpp is the de-facto on-device ASR runtime — a single self-contained binary, runs on CPU, zero dependencies. But Whisper is comparatively weak on Chinese. Now FunASR has a llama.cpp / GGUF runtime: the same download-and-run experience — one static binary, no Python, built-in VAD, any audio format — and it is about 2.7× more accurate than whisper.cpp on Chinese CPU.
-👉 Just want the download? Prebuilt binaries for Linux/macOS/Windows plus a one-page quickstart and benchmarks are on the Run FunASR on CPU (llama.cpp) page.
+👉 Just want the download? v0.2.0 ships nine Linux, macOS, and Windows archives with SHA-256 values. Use the llama.cpp / GGUF deployment page for the download matrix, quickstart, and known limitations.
Run it in 3 steps (all verified)
# 1. Download a prebuilt binary (linux-x64 / linux-arm64 / macos-arm64 / windows-x64) -wget https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.1.9/funasr-llamacpp-linux-x64-vulkan.tar.gz +wget https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-x64-vulkan.tar.gz tar xzf funasr-llamacpp-linux-x64-vulkan.tar.gz # Windows Vulkan asset: funasr-llamacpp-windows-x64-vulkan.zip (run with --backend vulkan) @@ -90,7 +98,8 @@What you get
- Built-in FSMN-VAD: automatic long-audio segmentation (`--vad fsmn-vad.gguf`); the bare binary reaches reference accuracy with no Python front end.
- Any audio: wav / mp3 / flac, any sample rate / channels — resampled inside the binary. -
- Multi-platform prebuilts: Linux x64/arm64, macOS arm64, Windows x64, including Linux/Windows Vulkan and Windows CUDA acceleration packages (Release); one CMake command to build for others. +
- Multi-platform prebuilts: v0.2.0 ships nine Linux x64/arm64, macOS arm64, and Windows x64 archives, including Linux/Windows Vulkan and Windows CUDA packages (Release); use one CMake command to build for other targets. +
- Direct subtitle output: the Fun-ASR-Nano, SenseVoiceSmall, and Paraformer CLIs can emit timestamped SRT subtitles with
--srt.