Skip to content

voice.md

Latest commit

 

History

History
117 lines (82 loc) · 4.5 KB

voice.md

File metadata and controls

117 lines (82 loc) · 4.5 KB
Raw

Voice

Hold Alt-t in Neovim, ask aloud, let go: Pair answers under the line you asked on, and can answer aloud too. Hearing and speaking run on your CPU with LFM3-Omni, one model for both, by default; whisper.cpp or OpenRouter can hear instead. Voice is off until you set it up; this takes about five minutes and one 4.0 GB download.

1. Choose who hears you

  • LFM3-Omni, on your machine. Nothing leaves it. It hears well (2% of words wrong on human speech) and reads answers aloud in a natural voice. It needs a 4.0 GB download that Pair makes only when you say so, 8.2 GB of disk in all, and up to 6 GB of memory while it speaks. It needs Python 3.12 to 3.14; Pair installs the rest itself.

  • whisper.cpp, on your machine. Smaller (142 MB) and lighter, and it reads nothing aloud by itself. Build it once:

    git clone --depth 1 https://github.com/ggml-org/whisper.cpp ~/src/whisper.cpp
cd ~/src/whisper.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF
cmake --build build -j --target whisper-cli
sh models/download-ggml-model.sh base.en

  • OpenRouter. No build: a transcription model hears the recording, and a speech model can read answers aloud. The recording leaves your machine, and each question costs one transcription. You need an OpenRouter key.

You also need a recorder: pw-record, parec or arecord, which most desktops have.

2. Add it to the config

For LFM3-Omni:

"voice": {}

Then, in Neovim, install it once:

:Pair voice on

Neovim says what comes and how big before it starts (Voice: installing … voice's Python packages (42 MB from PyPI, …), then LFM3-Omni-2B, 4.0 GB from Hugging Face …), then how far it is. It takes a minute or two on a fast line. The packages go into the same venv as search by meaning's, by their pinned SHA-256. :Pair voice off stops it, or deletes the model later.

For whisper.cpp:

"voice": {
  "model": "~/src/whisper.cpp/models/ggml-base.en.bin",
  "whisper": "~/src/whisper.cpp/build/bin/whisper-cli"
}

For OpenRouter (the key in a file, or apiKeyEnv, or auth after pair auth login openrouter):

"voice": {
  "openrouter": {
    "apiKeyFile": "~/.config/openrouter.key",
    "hear": "openai/whisper-large-v3-turbo",
    "speak": {"model": "x-ai/grok-voice-tts-1.0", "voice": "eve"}
  }
}

Short answers are read aloud only when you are away from the editor: LFM3-Omni does it by default, OpenRouter with speak; with whisper.cpp, only when you set speak too ("speak": "omni" lets LFM3-Omni read, or a speech command).

3. Check it

pair doctor

The voice line says what records and what hears: LFM3-Omni and whisper.cpp hear a second of silence and say how long it took, OpenRouter is asked whether the key and the models are taken. It records nothing, sends no audio and downloads nothing; without LFM3-Omni's files yet, it says to run :Pair voice on.

4. Ask

  1. In Neovim, on the line you want to ask about, hold Alt-t. A meter shows under the line: listening let go to ask.

    Listening: the meter under the line

  2. Say your question, and let go. The line says hearing, then the words Pair heard, then its answer in their place, with M-j when it points at a line. The whole answer is in the chat too.

    The answer under the line you asked on; the line it points at has its note

Hold Alt-t: a meter under your line; let go: the words Pair heard, then its answer in their place

The first question after the download takes about 15 s to hear, while LFM3-Omni packs its weights for your CPU, once. From then on a short question is heard in about a second and a half.

If your terminal makes holding awkward, tap Alt-t instead: Pair listens until you press it again, and the line says M-t again to ask. Or set listen_mode = 'toggle' in the plugin's setup: press to start, press again to ask. :Pair listen works without the key.

5. Talk with it

To hear every answer, not only when you are away, add "talk": true:

"voice": {"talk": true}

Ask aloud, and Pair answers aloud under your line: the first words come a few seconds after the answer. Ask the next question the same way. Start typing, or press Alt-t again, and it stops at once. Only answers to what you asked are ever spoken; Pair's own remarks stay quiet text, as rare as before.

More

Push-to-talk: every field, the key's timing, spoken answers, and what goes where.