Practical pattern: 1) maintain voices.json mapping locale -> voiceId + version; 2) generate one TTS file per locale+script and cache by content hash to avoid duplicates; filename: {campaign}_{locale}_v{voiceVersion}_vid{NN}.wav; 3) throttle concurrency (4–8), retry with exponential backoff for rate limits; 4) cut bitrate/trim silence and reuse identical clips across videos to save cost; 5) store artifacts (S3/GCS) and reference in CI render step.

For API details, see ElevenLabs docs.