diff options
| author | Somhairle H. Marisol <[email protected]> | 2026-09-29 10:54:22 +0800 |
|---|---|---|
| committer | Somhairle H. Marisol <[email protected]> | 2026-09-29 10:54:22 +0800 |
| commit | b927b2c2746efaf7f91d03cec1c123a3d52bf62b (patch) | |
| tree | 5c08f688691a4f4e8472759258c3021a46be9669 /THIRD_PARTY_NOTICES.md | |
| parent | 60f05cebd8e7968621a0ce015cf72987ec7577c7 (diff) | |
| download | mic-clipper-b927b2c2746efaf7f91d03cec1c123a3d52bf62b.tar.gz | |
fix(vad): bundle local Silero model
[问题原因]
- 先前实现会在包内权重缺失时读取外部 faster-whisper 资源,无法满足自包含部署要求。
[问题根因]
- 项目包中尚未包含已验证的 ONNX 权重,VAD 路径保留了开发期回退逻辑。
[修复方案]
- 提交已校验 SHA-256 的本地 Silero ONNX 权重并移除外部包回退。
- 增加模型打包和缺失模型失败路径测试,并更新来源、MIT 许可和哈希说明。
[影响范围]
- VAD 现在只加载项目包内模型,不联网且不依赖 faster-whisper 运行时资源。
- 未安装、启用或启动任何用户 systemd 服务。
Diffstat (limited to 'THIRD_PARTY_NOTICES.md')
| -rw-r--r-- | THIRD_PARTY_NOTICES.md | 14 |
1 files changed, 8 insertions, 6 deletions
diff --git a/THIRD_PARTY_NOTICES.md b/THIRD_PARTY_NOTICES.md index e38fc8e..e147b4e 100644 --- a/THIRD_PARTY_NOTICES.md +++ b/THIRD_PARTY_NOTICES.md @@ -2,12 +2,14 @@ ## Silero VAD ONNX weight -The intended bundled file is `src/mic_clipper/assets/silero_vad_v6.onnx`. +The bundled file is `src/mic_clipper/assets/silero_vad_v6.onnx`. -Its local source is the `assets/silero_vad_v6.onnx` file in the verified -`faster-whisper` 1.2.1 installation. That package's installed metadata declares -the MIT license and its VAD implementation attributes the model integration to +It was copied from `assets/silero_vad_v6.onnx` in the locally verified +`faster-whisper` 1.2.1 installation. Both files have SHA-256 +`4cbf549b8326f60f80f2536d9eefeb450a9abe83365a098031c89719f1be17d2`. +The installed package metadata declares the MIT license, and its VAD +implementation attributes the model integration to [`snakers4/silero-vad`](https://github.com/snakers4/silero-vad). -The ONNX weight has no network retrieval path in this project. The asset must be -present before deployment; it is intentionally not downloaded at runtime. +The ONNX weight is bundled and has no network retrieval path. The application +never downloads models at runtime. |
