Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
<!DOCTYPE html>
<html lang="zh">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>在 CPU 上运行 FunASR:llama.cpp / GGUF 运行时(whisper.cpp 中文替代)| FunASR</title>
<meta name="description" content="下载即用的 FunASR llama.cpp 运行时:单个二进制、无需 Python/GPU,在 CPU/边缘端跑 SenseVoice/Paraformer/Fun-ASR-Nano,内置 VAD,中文 CER 约为 whisper.cpp 的 1/3。Linux/macOS/Windows 预编译二进制。">
<meta name="keywords" content="FunASR llama.cpp, GGUF 语音识别, CPU 语音识别, 离线语音识别, whisper.cpp 替代, 端侧 ASR, 无 GPU 语音转文字, SenseVoice GGUF, 中文语音识别">
<link rel="icon" href="/logo.png">
<meta property="og:title" content="在 CPU 上运行 FunASR:llama.cpp / GGUF 运行时(whisper.cpp 中文替代)| FunASR"><meta property="og:description" content="下载即用的 FunASR llama.cpp 运行时:单个二进制、无需 Python/GPU,在 CPU/边缘端跑 SenseVoice/Paraformer/Fun-ASR-Nano,内置 VAD,中文 CER 约为 whisper.cpp 的 1/3。Linux/macOS/Windows 预编译二进制。">
<meta property="og:type" content="website"><meta property="og:url" content="https://www.funasr.com/llama-cpp.html">
<meta property="og:image" content="https://www.funasr.com/logo.png">
<link rel="canonical" href="https://www.funasr.com/llama-cpp.html">
<link rel="alternate" hreflang="en" href="https://www.funasr.com/en/llama-cpp.html"><link rel="alternate" hreflang="zh" href="https://www.funasr.com/llama-cpp.html">
<link rel="alternate" hreflang="x-default" href="https://www.funasr.com/en/llama-cpp.html">
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap" rel="stylesheet">
<script type="application/ld+json">
{"@context":"https://schema.org","@type":"SoftwareApplication","name":"FunASR llama.cpp runtime","description":"下载即用的 FunASR llama.cpp 运行时:单个二进制、无需 Python/GPU,在 CPU/边缘端跑 SenseVoice/Paraformer/Fun-ASR-Nano,内置 VAD,中文 CER 约为 whisper.cpp 的 1/3。Linux/macOS/Windows 预编译二进制。","applicationCategory":"DeveloperApplication","operatingSystem":"Linux, macOS, Windows","offers":{"@type":"Offer","price":"0","priceCurrency":"USD"},"author":{"@type":"Organization","name":"FunASR / Tongyi Lab"},"url":"https://www.funasr.com/llama-cpp.html"}
</script>
<style>:root{--primary:#2563eb;--primary-dark:#1d4ed8;--bg:#fff;--surface:#f8fafc;--border:#e2e8f0;--text:#0f172a;--text-soft:#475569;--text-muted:#94a3b8;--radius:12px}
*{margin:0;padding:0;box-sizing:border-box}
body{font-family:'Inter',-apple-system,sans-serif;color:var(--text);line-height:1.7}
a{color:var(--primary);text-decoration:none}a:hover{text-decoration:underline}
.container{max-width:800px;margin:0 auto;padding:0 24px}
.nav{position:fixed;top:0;left:0;right:0;z-index:100;background:rgba(255,255,255,0.9);backdrop-filter:blur(12px);border-bottom:1px solid var(--border);padding:14px 0}
.nav .container{max-width:1120px;display:flex;align-items:center;gap:24px}
.nav-logo{font-size:1.2rem;font-weight:800;color:var(--text)}.nav-logo span{color:var(--primary)}
.nav-links{display:flex;gap:20px;margin-left:auto;flex-wrap:wrap}
.nav-links a{font-size:0.85rem;font-weight:500;color:var(--text-soft)}
.nav-btn{padding:7px 16px;background:var(--primary);color:#fff;border-radius:7px;font-size:0.83rem;font-weight:600}
article{padding:120px 0 80px}
h1{font-size:2rem;font-weight:800;line-height:1.3;margin-bottom:14px}
h2{font-size:1.3rem;font-weight:700;margin:36px 0 14px;padding-top:16px;border-top:1px solid var(--border)}
h3{font-size:1.05rem;font-weight:700;margin:22px 0 10px}
p{margin-bottom:14px;color:var(--text-soft)}
ul,ol{margin:0 0 14px 22px;color:var(--text-soft)}li{margin-bottom:6px}
table{width:100%;border-collapse:collapse;margin:16px 0;font-size:0.86rem}
th,td{padding:10px 12px;text-align:left;border-bottom:1px solid var(--border)}
th{font-weight:600;background:var(--surface);font-size:0.78rem;color:var(--text-muted)}
pre{background:#1e293b;color:#e2e8f0;padding:18px;border-radius:10px;overflow-x:auto;font-size:0.82rem;line-height:1.7;margin:14px 0}
code{font-family:ui-monospace,Menlo,monospace}
p code,li code{background:var(--surface);padding:2px 6px;border-radius:5px;font-size:0.85em;color:var(--primary-dark)}
.post-meta{color:var(--text-muted);font-size:0.85rem;margin-bottom:24px}
.cta{background:var(--surface);border:1px solid var(--border);border-radius:var(--radius);padding:24px;margin:28px 0;text-align:center}
.cta a.btn{display:inline-block;padding:10px 24px;background:var(--primary);color:#fff;border-radius:8px;font-weight:600}
.post-list{list-style:none;margin:0;padding:0}
.post-list li{margin:0 0 8px}
.post-card{display:block;border:1px solid var(--border);border-radius:var(--radius);padding:20px 22px;margin-bottom:14px}
.post-card:hover{border-color:var(--primary);text-decoration:none}
.post-card h2{border:none;padding:0;margin:0 0 6px;font-size:1.15rem;color:var(--text)}
.post-card p{margin:0;color:var(--text-soft);font-size:0.9rem}
.post-card .date{color:var(--text-muted);font-size:0.8rem}
footer{background:#0f172a;color:#94a3b8;padding:28px 0;text-align:center;font-size:0.8rem}
footer a{color:#94a3b8}
.dl-grid{display:grid;grid-template-columns:repeat(auto-fit,minmax(200px,1fr));gap:12px;margin:18px 0}
.dl-card{border:1px solid var(--border);border-radius:10px;padding:16px 18px}
.dl-card b{color:var(--text)}
.hero{text-align:center;padding:20px 0 8px}
.hero .tags{color:var(--text-muted);font-size:0.9rem;margin-top:8px}
@media(max-width:900px){.nav-links{display:none}.nav .container{gap:16px}.nav-logo{margin-right:auto}h1{font-size:1.75rem}table{table-layout:fixed;font-size:0.78rem}th,td{padding:8px 4px;overflow-wrap:anywhere}p code,li code{overflow-wrap:anywhere;word-break:break-word}}
</style>
</head>
<body>
<nav class="nav"><div class="container"><a href="/" class="nav-logo">Fun<span>ASR</span></a><div class="nav-links"><a href="/#capabilities">核心能力</a><a href="/#usage">如何使用</a><a href="/#benchmark">性能</a><a href="/#demo">演示</a><a href="https://modelscope.github.io/FunASR/zh/" target="_blank">文档</a><a href="/quickstart.html">快速上手</a><a href="/models.html">模型</a><a href="/ecosystem.html">生态</a><a href="/blog/">技术博客</a><a href="/vs-whisper.html">对比 Whisper</a><a href="/donors.html">功德榜</a></div><a href="https://www.funasr.com/en/llama-cpp.html" style="font-size:0.83rem;font-weight:500;color:var(--text-soft)">EN</a><a href="https://github.com/modelscope/FunASR" target="_blank" class="nav-btn">GitHub ★</a></div></nav>
<article><div class="container">
<div class="hero"><h1>在 CPU 上运行 FunASR —— llama.cpp / GGUF 运行时</h1>
<p class="tags">单个二进制 · 无需 Python · 无需 GPU · 量化权重 · whisper.cpp 式的端侧语音识别,中文更强</p></div>
<p><strong>FunASR on llama.cpp</strong> 之于 FunASR,正如 <a href="https://github.com/ggml-org/whisper.cpp" target="_blank">whisper.cpp</a> 之于 Whisper:把 SenseVoice / Paraformer / Fun-ASR-Nano 三个模型搬到 <a href="https://github.com/ggml-org/llama.cpp" target="_blank">ggml</a> 栈上,**在没有 GPU、没有 Python 的地方也能跑**(笔记本、边缘盒子、嵌入式 C/C++ 应用),与 GPU 上的 PyTorch / vLLM 路线互补。VAD(FSMN)已内置进二进制。</p>
<h2>下载预编译二进制(下载即用)</h2>
<p>覆盖 Linux(x64/arm64)、macOS(arm64)、Windows(x64),静态自包含、无依赖。三个仓库的 Releases 都能下:</p>
<div class="dl-grid">
<a class="dl-card" href="https://github.com/modelscope/FunASR/releases" target="_blank"><b>⬇ modelscope/FunASR</b><br><span style="color:var(--text-soft);font-size:0.88rem">三模型全套 runtime</span></a>
<a class="dl-card" href="https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9" target="_blank"><b>⬇ 当前 runtime v0.1.8</b><br><span style="color:var(--text-soft);font-size:0.88rem">Linux/Windows Vulkan + Windows CUDA + 9 个平台包</span></a>
<a class="dl-card" href="https://github.com/modelscope/FunASR/releases/tag/v1.3.26" target="_blank"><b>Python 包 v1.3.26</b><br><span style="color:var(--text-soft);font-size:0.88rem">funasr-server --hub ms 修复 + runtime 下载表</span></a>
<a class="dl-card" href="https://github.com/QwenAudio/SenseVoice/releases" target="_blank"><b>⬇ QwenAudio/SenseVoice</b><br><span style="color:var(--text-soft);font-size:0.88rem">多语种 + 情感/事件</span></a>
<a class="dl-card" href="https://github.com/QwenAudio/Fun-ASR/releases" target="_blank"><b>⬇ QwenAudio/Fun-ASR</b><br><span style="color:var(--text-soft);font-size:0.88rem">Fun-ASR-Nano(LLM-ASR)</span></a>
</div>
<p>当前推荐下载入口:<a href="https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9" target="_blank">runtime-llamacpp-v0.1.9</a>。其中常用加速包包括 <a href="https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.1.9/funasr-llamacpp-linux-x64-vulkan.tar.gz" target="_blank">funasr-llamacpp-linux-x64-vulkan.tar.gz</a>、<a href="https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.1.9/funasr-llamacpp-windows-x64-vulkan.zip" target="_blank">funasr-llamacpp-windows-x64-vulkan.zip</a> 和 <a href="https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.1.9/funasr-llamacpp-windows-x64-cuda.zip" target="_blank">funasr-llamacpp-windows-x64-cuda.zip</a>。Python 用户请安装 <code>python -m pip install -U "funasr==1.3.26"</code>;<a href="https://github.com/modelscope/FunASR/releases/tag/v1.3.26" target="_blank">v1.3.26 Latest Release</a> 也列出了同一组 runtime 下载链接和 SHA-256。</p>
<h2>三行跑通</h2>
<pre># 1) 解压下载的二进制,拉一个模型(自动下 GGUF + VAD)
bash download-funasr-model.sh sensevoice ./gguf
# 2) 直接出中文(二进制内置 detok,无需 Python)
llama-funasr-sensevoice -m ./gguf/SenseVoiceSmall-f16.gguf --vad ./gguf/fsmn-vad.gguf -a audio.wav
# -> 欢迎大家来体验达摩院推出的语音识别模型</pre>
<p>换模型:<code>download-funasr-model.sh paraformer</code> 配 <code>llama-funasr-paraformer</code>;<code>download-funasr-model.sh nano</code> 配 <code>llama-funasr-cli</code>(Fun-ASR-Nano,中英日 + 中文方言/口音)。31 语种属于独立的 MLT-Nano checkpoint。</p>
<h2>精度:中文上对 whisper.cpp 是碾压</h2>
<p>同一 184 集中文测试集,字错误率 CER(micro-avg,越低越准):</p>
<table>
<tr><th>模型</th><th>CER ↓</th><th>说明</th></tr>
<tr><td><strong>FunASR SenseVoice</strong></td><td><strong>8.01%</strong></td><td>多语种 + 情感/事件</td></tr>
<tr><td><strong>FunASR Paraformer</strong></td><td><strong>9.85%</strong></td><td>非自回归,工业级中文</td></tr>
<tr><td><strong>FunASR Fun-ASR-Nano</strong></td><td><strong>8.30%</strong></td><td>LLM-ASR,中英日 + 中文方言/口音</td></tr>
<tr><td>whisper.cpp small</td><td>22.12%</td><td></td></tr>
<tr><td>whisper.cpp large-v3-turbo</td><td>23.15%</td><td></td></tr>
<tr><td>whisper.cpp base</td><td>31.33%</td><td></td></tr>
</table>
<p>FunASR 的中文 CER 约为 whisper.cpp 的 1/3。完整方法学见各仓 <a href="https://github.com/modelscope/FunASR/blob/main/runtime/llama.cpp/BENCHMARKS.md" target="_blank">runtime/llama.cpp/BENCHMARKS.md</a>。</p>
<h2>模型体积与精度(量化档)</h2>
<p>量化后体积更小、速度更快,精度几乎不变(全 184 集 micro-CER):</p>
<table>
<tr><th>模型</th><th>档</th><th>体积</th><th>CER ↓</th><th>速度</th></tr>
<tr><td>SenseVoice encoder</td><td>f16</td><td>470 MB</td><td>8.01%</td><td>23×</td></tr>
<tr><td>SenseVoice encoder</td><td><strong>q8</strong></td><td><strong>254 MB</strong></td><td>7.99%</td><td>27×</td></tr>
<tr><td>Paraformer encoder</td><td>f16</td><td>435 MB</td><td>9.85%</td><td>22×</td></tr>
<tr><td>Paraformer encoder</td><td><strong>q8</strong></td><td><strong>237 MB</strong></td><td>9.78%</td><td>27×</td></tr>
<tr><td>Fun-ASR-Nano LLM*</td><td>q8_0</td><td>805 MB</td><td>8.30%</td><td>6.0×</td></tr>
<tr><td>Fun-ASR-Nano LLM*</td><td><strong>q5_K_M</strong></td><td>551 MB</td><td><strong>8.25%</strong></td><td>5.7×</td></tr>
<tr><td>Fun-ASR-Nano LLM*</td><td><strong>q4_K_M</strong></td><td><strong>484 MB</strong></td><td>8.35%</td><td>6.1×</td></tr>
</table>
<p style="font-size:0.9rem">* Fun-ASR-Nano 另需 encoder(f16 470 MB)。q8 编码器体积减半、CER 反而略降;Nano LLM 三档精度几乎无差——推荐 <strong>q4_K_M</strong>(最小)或 <strong>q5_K_M</strong>(最佳)。</p>
<h2>包含什么</h2>
<ul>
<li>6 个二进制:<code>llama-funasr-{cli,encoder,embd,sensevoice,paraformer,vad}</code>,静态、无 .so 依赖。</li>
<li>内置 FSMN-VAD(<code>--vad</code>)、二进制内 detok 直接出文字、kaldi 兼容 fbank 前端。</li>
<li>GGUF 模型在 Hugging Face:<a href="https://huggingface.co/FunAudioLLM" target="_blank">FunAudioLLM</a> / <a href="https://huggingface.co/funasr" target="_blank">funasr</a>(f16 / f32 / <strong>q8</strong> + 嵌入 vocab;<strong>q8 体积约为 f16 的一半、精度不变</strong>,下载更快)。</li>
<li>源码与文档:<a href="https://github.com/modelscope/FunASR/tree/main/runtime/llama.cpp" target="_blank">runtime/llama.cpp/</a>(README / DESIGN / BENCHMARKS)。</li>
</ul>
<div class="cta"><p><strong>觉得有用就给个 Star 支持一下 👇</strong> 全家桶开源、可商用。</p>
<p><a class="btn" href="https://github.com/modelscope/FunASR" target="_blank">⭐ Star FunASR</a></p>
<p style="margin-top:10px;font-size:0.85rem">也欢迎 Star:<a href="https://github.com/QwenAudio/SenseVoice" target="_blank">SenseVoice</a> · <a href="https://github.com/QwenAudio/Fun-ASR" target="_blank">Fun-ASR</a> · <a href="https://github.com/modelscope/FunClip" target="_blank">FunClip</a></p></div>
<p style="font-size:0.9rem">延伸阅读:<a href="/blog/funasr-llama-cpp-whisper-cpp-alternative.html">FunASR 跑进 llama.cpp(whisper.cpp 替代)详解</a></p>
</div></article>
<footer><div class="container">© 2024-2026 Tongyi Lab, Alibaba Group · <a href="https://github.com/modelscope/FunASR/blob/main/LICENSE">MIT License</a></div></footer>
</body>
</html>