opencode-read-aloud

Reads OpenCode's last answer aloud and highlights each word as it is spoken. Press ctrl+s in a session to start. Press it again to pause or resume.

The same 20 seconds with sound: assets/demo.mp4.

• The sentence being read gets a faint tint and the current word a stronger one. The transcript scrolls to keep that paragraph on screen.
• Only the answer is read: the text after the last tool call of the latest assistant message. Code blocks, tables and rules are skipped.
• Option+click (alt+click) any word of an answer to start reading from that word. This works for older answers too.
• The controls on the prompt row are clickable: play/pause, slower -, faster +, and stop. On terminals narrower than 80 columns only play/pause and the speed are shown.
• Speed changes keep the pitch and are remembered between sessions. The default is 1.7x.
• On macOS the AirPods button and the keyboard's play/pause key pause and resume reading while it is active, instead of starting your music. This needs the Xcode command line tools (xcode-select --install), which compile a small helper on first use. Once reading stops, the keys go back to your music app.
• Audio is cached, so replaying an answer or jumping around in it costs nothing after the first read.

• A Speechify or ElevenLabs API key. Both report when each word is spoken, which is what drives the highlighting. When both keys are set, Speechify is used.
• ffmpeg on your PATH (brew install ffmpeg). It handles seeking and speed changes.

Set a key with an environment variable or a key file:

export SPEECHIFY_API_KEY=sk_...
# or
echo sk_... > ~/.local/share/opencode/speechify.key

export ELEVENLABS_API_KEY=sk_...
# or
echo sk_... > ~/.local/share/opencode/elevenlabs.key

The voice and model can be changed with the speechify and elevenlabs options, or with these environment variables:

• Speechify: SPEECHIFY_VOICE_ID (default Dominic, dominic_32) and SPEECHIFY_MODEL_ID (default simba-3.2, English). GET https://api.speechify.ai/v1/voices lists the others; geffen_32 is a female voice made for the same model.
• ElevenLabs: ELEVENLABS_VOICE_ID (default George, JBFqnCBsd6RMkjVDRZzb) and ELEVENLABS_MODEL_ID (default eleven_flash_v2_5).

opencode plugin opencode-read-aloud -g

That installs the package into your global OpenCode config and adds it to tui.json. Drop -g for the current project only. Restart OpenCode afterwards; plugins load at startup.

git clone https://github.com/moritzWa/opencode-read-aloud ~/.config/opencode/tui-plugins/read-aloud

Then add it to ~/.config/opencode/tui.json:

{
  "plugin": ["./tui-plugins/read-aloud/index.tsx"]
}

OpenCode loads the TypeScript source directly, so there is nothing to build.

{
  "plugin": [
    [
      "opencode-read-aloud",
      {
        "keybinds": { "speech_toggle": "ctrl+s", "speech_faster": "alt+=", "speech_slower": "alt+-" },
        "speechify": { "voice": "hugh_32" }
      }
    ]
  ]
}

All five commands are also in the command palette (ctrl+p) under "Speech".

• Finding the text. The plugin walks the rendered transcript for the markdown view whose content matches each answer part. It reads that view's prose blocks: paragraphs, headings and list items.
• Building the script. The prose is turned into a plain-text script, with a map from every script character back to its position in the markdown.
• Speech. The script goes to Speechify's or ElevenLabs' streaming with-timestamps endpoint. Playback starts on the first chunk. ElevenLabs times every character; Speechify times every word, and each character takes the start of its word.
• Audio. Audio is piped through ffmpeg (for the seek and the atempo speed change) into opentui's audio output.
• Highlighting. Every 40 ms the playback position is mapped to a word. That word and its sentence are highlighted by extending the block's onHighlight, using two styles registered on the block's syntax style and tinted from the theme's primary colour.
• Cache. Audio and timings are cached in ~/.cache/opencode/speech/, keyed by a hash of the voice, the model and the text. A cached reading from either provider is reused before anything new is requested.

MIT