Generate the voiceover

Every video gets a spoken voiceover by default — it's the single biggest lift in perceived quality. You turn the narration script into audio with HeyGen's text-to-speech, then hand the audio URL to the composition.

Skip this step only if the user's prompt explicitly asks for no narration — "no voiceover", "silent", "ambient only", "music only", "text only". When you skip, leave voiceover_url empty and tell the user the video is silent.

Inputs

• The narration script from generate-script (a short spoken line or two, sized to the starter's duration — distinct from the on-screen copy).
• The chosen composition's manifest.json, which carries a recommended_voice_id (the default voice for that starter).

Steps

1. Pick a voice. Use the starter's recommended_voice_id unless the user's prompt gives tone direction — then override from the catalog below.

2. Synthesize with the TTS helper:

   set +x
   RESULT=$(python3 /.agents/workspace/scripts/tts_client.py \
     "<narration script>" "<voice_id>")
   AUDIO_URL=$(echo "$RESULT" | jq -r '.audio_url')
   DURATION=$(echo "$RESULT" | jq -r '.duration')
   

   tts_client.py calls POST /v3/voices/speech and prints {audio_url, duration, voice_id, word_timestamps}. The sandbox proxy injects the HeyGen key — you pass no credential. (Dev test: prefix with HEYGEN_API_URL=https://api.dev.heygen.com.)

3. Check the duration against the starter. Compare DURATION to the starter's duration_seconds (from its manifest). The audio must fit:

   • DURATION ≤ starter duration → good, use it.
   • DURATION > starter duration → do not truncate. Trim the narration script (fewer words / shorter sentences) and re-synthesize until it fits, with ~0.3s of headroom. A tight, punchy line beats a rushed or cut-off one.

4. Save the audio URL for customize-composition:

   echo -n "$AUDIO_URL" > /.agents/workspace/output/voiceover_url.txt
   

   (If you skipped narration, write an empty file or skip this — customize will strip the audio element.)

Voice catalog (starfish engine)

Pick by tone. Tones are a starting point — the per-starter recommended_voice_id is the default; override only on explicit prompt direction.

Voice	voice_id	Tone — use for
Luna Bright (f)	2b0124724e144eedba61ecb590d6c266	Bright, friendly — consumer apps, social promos
Smooth Dev (m)	07d2ba65847541feb97abc9b60181555	Smooth, confident — product trailers, narration
Marcus – Professional (m)	0f50a7a5577e4cd583ba738094956899	Warm, clear, professional — explainers, how-it-works, B2B
Peaceful Paige (f)	035ec8f52ec247cdbed5f3bf9f7991ef	Calm, soft, soothing — wellness, meditation, premium calm
Energetic Mikey (m)	25a40947b46942c1a8e5bd329074bba2	Upbeat, punchy, high-energy — launches, hype, social

Override cues: "calm / relaxing / meditative / soothing" → Peaceful Paige · "energetic / exciting / hype / launch / punchy" → Energetic Mikey · "friendly / fun / approachable" → Luna Bright · "professional / corporate / clear" → Marcus · "smooth / cinematic / confident" → Smooth Dev.

(These are public starfish-engine voices; if a voice_id ever fails, list current options with GET /v3/voices?engine=starfish.)

Output of this step

• /.agents/workspace/output/voiceover_url.txt — the synthesized audio URL.

The renderer reads the <audio> src as a static HTML attribute, so this URL isn't a runtime variable — customize-composition bakes it into the staged composition's HTML (via apply_voiceover.py). Then move to customize-composition.