@spences10/pi-talk

Opt-in hold-to-talk dictation for my-pi and Pi. Hold Space to record, then release it to transcribe into the active editor. A quick Space tap remains a normal space.

Talk is off by default. It does not intercept Space until you enable it through /talk setup or /talk on.

Requirements

• Linux with PipeWire's pw-record
• Node.js 24.15 or newer
• Deepgram API key
• Pi 0.85.0 or newer

Installation

pi install npm:@spences10/pi-talk

Set up Talk inside Pi:

/talk setup

The setup flow explains where the key is stored, then opens a masked input. After setup, hold Space to dictate and release it to transcribe. The transcript stays in the editor for review and is never submitted automatically.

Commands

/talk setup   Store a Deepgram key and enable Talk
/talk on      Enable Talk with the stored or environment key
/talk off     Disable Talk without deleting the stored key
/talk forget  Disable Talk and delete the stored key
/talk status  Show state, key source, and the next action

Talk has three user-visible states:

• not configured: no stored or environment key; Space is untouched
• off: a key is available, but Space is untouched
• ready: hold-Space detection is active

Secret storage

/talk setup stores the key under packages.pi-talk in ~/.pi/agent/my-pi-secrets.json, or in the configured Pi agent directory. The file uses private permissions (0600). Writes preserve secrets owned by other packages.

The masked setup input does not add the key to the conversation or Pi session history. Talk notifications and errors never include the key.

You can use DEEPGRAM_API_KEY or DEEPGRAM instead of storing a key. An environment key does not enable Talk automatically; run /talk on.

Runtime behavior

• A normal Space tap passes through immediately, including when the next key is pressed before Space is released.
• Other typing, cursor movement, or pasted input cancels a pending hold. Fast typing never restores an earlier copy of the editor.
• Spaces inserted before a hold is recognized stay in the editor; further repeats are consumed while recording.
• A held Space starts mono 16-bit, 16 kHz capture through pw-record.
• Audio streams to Deepgram's live nova-3 API.
• Releasing Space finalizes the stream and inserts final transcript text at the end of the current editor contents. Typing, edits, and deletions made while waiting are preserved.
• Talk stops its recorder and socket when disabled or when the session shuts down.
• Audio and transcripts are not written to disk.

Terminals that send explicit Space repeat/release events use those signals to detect a hold. In WezTerm's default mode and other terminals without those events, Talk uses consecutive spaces from keyboard repeat: at least three Space events over 300 ms, allowing up to 650 ms for the first repeat and 180 ms between subsequent repeats. A pause of 180 ms after repeats stops recording. Typing another key cancels the sequence immediately; two Space taps are preserved as ordinary input.

Without release events, repeated taps with the same timing as keyboard repeat remain indistinguishable from a hold. Talk preserves existing editor text even if such a sequence activates recording.

In my-pi, --no-talk prevents the built-in extension from loading for one launch.

Local development

pnpm --filter @spences10/pi-talk run build
pi install ./packages/pi-talk
# or for one run only
pi -e ./packages/pi-talk

Using from a custom harness

import talk from '@spences10/pi-talk';

// Pass talk as an ExtensionFactory to your Pi runtime.

Development

Package scripts build transitive workspace dependencies first, then run local tools through Vite+ with vp exec.

pnpm --filter @spences10/pi-talk run check
pnpm --filter @spences10/pi-talk run test
pnpm --filter @spences10/pi-talk run build

License

MIT