Offline ASR and Wake-Word Engineering for a Local AI Controller
A Juejin engineering account separates on-demand speech recognition from always-on wake-word detection in a local AI controller. The author reports fixing Vosk model-selection order, adapting openWakeWord training for domestic network conditions, and comparing openWakeWord, Porcupine, and Vosk keyword detection; self-reported tests included a 2.06 ms mean inference time and 556 passing regression assertions. The author also says Chinese-accent wake rates and FunASR and faster-whisper accuracy were not measured, and the reported results have not been independently verified.
The account provides a concrete implementation path and measured tradeoffs for building offline, always-on voice interaction in environments where external training services or cloud inference are impractical.
一个本地 AI 中控要做到「随叫随到」,得先解决两件事:听清(ASR)和听见(唤醒)。 这两件事在国内网络环境下各有一串坑,我把踩过的都记在这了。
一、先把两件事拆开:很多人一开始就把它们混在一起
做语音时最常见的第一个错误,是把「语音识别」和「语音唤醒」当成一件事。
| 语音识别(ASR) | 语音唤醒(Wake Word) | |
|---|---|---|
| 干什么 | 把波形转成文字 | 判断「刚才那声是不是在叫我」 |
| 输出 | 一整句文本 | (keyword, score) 一个事件 |
| 谁来打分 | 语言模型(解码整段声学特征) | 小分类器(跑在冻结的特征提取器上) |
| 延迟要求 | 秒级可接受 | 毫秒级,必须常驻 |
| 能不能云端 | 可以(但违数据主权) | 绝对不行,要 7×24 在线 |
| 本项目对应 | Vosk / FunASR / faster-whisper | openWakeWord / Porcupine / Vosk 关键词 |
#openWakeWord#Porcupine#Vosk#FunASR#faster-whisper#Picovoice#Baidu AI Studio#ModelScope#ONNX Runtime#SAPI5