跳转至

type: concept tags: [audio, mlx, on-device, speech-to-text, macos, apple-silicon, 其他] related: [[gemma4-ondevice]], , [[gemma4-audio-mlx]], sources: - https://simonwillison.net/2026/Apr/13/gemma-4-audio-with-mlx/ created: 2026-04-14


Gemma 4 端侧音频处理(MLX)

在 Apple Silicon 设备上使用 Gemma 4 E2B 模型进行本地音频转录的方案。

技术实现

通过 框架在 macOS 上本地运行 Gemma 4 E2B 模型(10.28GB):

uv run --python 3.13 --with mlx_vlm --with torchvision --with gradio \
  mlx_vlm.generate \
  --model google/gemma-4-e2b-it \
  --audio file.wav \
  --prompt "Transcribe this audio" \
  --max-tokens 500 --temperature 1.0

实际效果

  • 在 14 秒 WAV 文件上测试,转录质量良好
  • 存在少量误识别(如 "This right here" 被识别为 "This front here")
  • 整体可用级别,适合离线转录场景

为什么重要

这是[[gemma4-audio-mlx]]的一个重要里程碑: 1. 全离线:音频数据完全不离开设备 2. 多模态统一:同一模型处理视觉、文本和音频,无需 [[gemma4-audio-mlx]] 等专用模型 3. MLX 优化:Apple Silicon 原生推理框架,充分发挥 M 系列芯片性能 4. 开发者友好:一条 uv 命令即可运行,降低端侧 AI 使用门槛

对 而言,这展示了端侧多模态模型已从"实验室演示"进入"开发者可用"阶段。

核心问题

Computer use agents automate digital tasks by directly interacting with graphical user interfaces (GUIs) on computers and mobile devices, offering significant potential to enhance human productivity by completing an open-ended space of user queries. However, current agents face significant challenges: imprecise grounding of GUI elements, difficulties with long-horizon task planning, and performance bottlenecks from relying on single generalist models for diverse cognitive tasks. To this end, we introduce Agent S2, a novel compositional framework that delegates cognitive responsibilities across various generalist and specialist models. We propose a novel Mixture-of-Grounding technique to achieve precise GUI localization and introduce Proactive Hierarchical Planning, dynamically refining act

为什么重要

本研究/产品对手机端 AIOS 生态有重要参考价值。推动端侧 AI 从概念走向实际部署。

关联

  • [[clawmobile-agentic]] — Agent 系统架构
  • [[mnn-350]] — 推理引擎
  • [[kv-cache-quantization-ondevice]] — 内存优化