generate-cues

name	generate-cues
description	Pre-generate the TTS announcement clips ("Sample 1", "Sample 2") that tune-preset and audition-preset stitch into combined comparison files so the user can identify which take is which without context-switching to a UI. Uses edge-tts (Microsoft Edge neural voices) by default; falls back to espeak-ng if edge-tts isn't installed or has no network. Idempotent — only re-runs when forced.
disable-model-invocation	true
allowed-tools	Bash(mkdir *), Bash(test *), Bash(ls *), Bash(cat *), Bash(ffmpeg *), Bash(ffprobe *), Bash(espeak-ng *), Read, Write

Generate TTS Cues

Pre-render the spoken announcements that downstream A/B comparison tooling concatenates with audio clips. Generated once, reused on every tune session and audition.

Resolve paths

PLUGIN_DATA_DIR="${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/audio-production"
TTS_DIR="$PLUGIN_DATA_DIR/tts"
VENV="$PLUGIN_DATA_DIR/venv"

Inputs

$ARGUMENTS:

• --voice=<edge-voice-id> — default en-US-AndrewNeural (clear, neutral male). See python -m edge_tts --list-voices for the catalogue.
• --engine=<edge-tts|espeak-ng> — default edge-tts. If edge-tts isn't installed or the network is unreachable, fall back to espeak-ng.
• --force — regenerate even if the cue files already exist.
• --cues="Sample 1,Sample 2,Original,Variant A,Variant B" — comma-separated list of phrases to render. Default just Sample 1,Sample 2 (sufficient for tune-preset and audition-preset).

Procedure

1. Skip if already done

If <TTS_DIR>/sample-1.wav and <TTS_DIR>/sample-2.wav exist and --force was not passed, report "cues already present" and exit.

2. Pick the engine

if [ "<engine>" = "edge-tts" ] && "$VENV/bin/python" -c "import edge_tts" 2>/dev/null; then
  ENGINE=edge-tts
elif command -v espeak-ng >/dev/null; then
  ENGINE=espeak-ng
else
  echo "No TTS engine available. Run /audio-production:install-deps to install edge-tts."
  exit 1
fi

3. Render each cue

For each phrase in the cue list, derive a slug (lowercase, spaces → hyphens) and render to <TTS_DIR>/<slug>.wav at 48 kHz mono PCM (matching the rest of the plugin's audio).

edge-tts path

mkdir -p "$TTS_DIR"
"$VENV/bin/python" - <<'PY'
import asyncio, os, re, sys
import edge_tts

phrases = sys.argv[1].split(",")
voice = sys.argv[2]
out_dir = sys.argv[3]

async def render(text, path):
    comm = edge_tts.Communicate(text, voice)
    await comm.save(path)

async def main():
    for p in phrases:
        slug = re.sub(r"[^a-z0-9]+", "-", p.lower()).strip("-")
        await render(p.strip(), os.path.join(out_dir, f"{slug}.mp3"))

asyncio.run(main())
PY "<cues>" "<voice>" "$TTS_DIR"

Then transcode each MP3 to canonical 48 kHz mono PCM WAV and remove the MP3:

for f in "$TTS_DIR"/*.mp3; do
  ffmpeg -y -i "$f" -ac 1 -ar 48000 -c:a pcm_s16le "${f%.mp3}.wav"
  rm "$f"
done

espeak-ng path (fallback)

for phrase in <cue list>; do
  slug=...
  espeak-ng -v en+m3 -s 150 -w "$TTS_DIR/${slug}.raw.wav" "$phrase"
  ffmpeg -y -i "$TTS_DIR/${slug}.raw.wav" -ac 1 -ar 48000 -c:a pcm_s16le "$TTS_DIR/${slug}.wav"
  rm "$TTS_DIR/${slug}.raw.wav"
done

4. Pad each cue with silence

Add 200 ms of leading silence and 400 ms of trailing silence so the cue doesn't sit flush against the audio sample it announces:

ffmpeg -y -i "$TTS_DIR/<slug>.wav" \
  -af "adelay=200|200,apad=pad_dur=0.4" \
  -c:a pcm_s16le "$TTS_DIR/<slug>.padded.wav"
mv "$TTS_DIR/<slug>.padded.wav" "$TTS_DIR/<slug>.wav"

5. Report

List each cue file with duration and the engine/voice used. Suggest the user audition them once:

mpv ~/.local/share/claude-plugins/audio-production/tts/sample-1.wav

Notes

• edge-tts calls Microsoft's neural TTS endpoint over the network; expect ~1–2 seconds per phrase. No API key required, but offline use needs the espeak-ng fallback.
• Cues are 48 kHz mono PCM to match plugin canonical format — concatenation with audio clips is then a single ffmpeg -i a.wav -i b.wav -filter_complex concat call with no resampling.
• This skill is idempotent — re-run with --force if you change voices or want to regenerate.
• No cue content is sensitive; voicing Sample 1 to a Microsoft endpoint is acceptable for everyone.