Hey Hearthlen Opened YouTube Without a Cloud Model

说明与用途

Hearthlen 在 4.2 秒内本地打开 YouTube，使用本地唤醒词、语音识别和 LLM 模型，内存保持在 250 MB 以下。

问题

How to open YouTube locally without relying on a cloud language model while keeping latency low and memory usage within an 8 GB laptop budget.

结果

The Hearthlen assistant opens YouTube locally in 4.2 s, with 9 µs skill matching and 364 ms STT, staying within a 150–250 MB always‑on memory budget on an 8 GB laptop.

工作流

• 1. Wake‑word "hey Hearthlen" detected via pocketsphinx (15.6 MB binary).
• 2. Speech‑to‑text with fast‑whisper tiny.en (int8) loads once, costing 459 932 kB, then transcribes command in 364 ms.
• 3. match() function identifies skill "play video" in 9 µs.
• 4. Skill executes local YouTube search to resolve title to watch ID (4221 ms total).
• 5. Browser opens https://www.youtube.com/watch?v=JGwWNGJdvx8.
• 6. If command not a skill, 360 M 4‑bit LLaMA via llama.cpp replies in 758 ms and speaks the answer via Piper voice (2048 ms).

证据

"Play Shape of You on YouTube" opened the watch page in 4.2 seconds. Deciding that this was a YouTube request took 9 microseconds.

来源

Speech‑to‑text is faster‑whisper tiny.en at int8. On a clip of the YouTube sentence it took 364 ms once the model was already loaded, and the transcript was “Play Shape of you on YouTube.” Keeping that model loaded costs 459,932 kB in the bench process.

来源

The listener is a user service. Before the first utterance it was 15.6 MB. It is waiting for “hey Hearthlen” with pocketsphinx and a pronunciation written for that made‑up name.

来源

If the sentence is not a skill, a 360M model (4‑bit, through llama.cpp) answers and the process exits. One of those replies took 758 ms: “What do you call a person with no friends? A loner.”

来源