elevenlabs

Create an ElevenLabs client with your API key and call client.text_to_speech.convert with text, a voice_id, and a model like eleven_multilingual_v2. Save the returned audio stream to an MP3 file using the save helper or by writing chunks.

Use eleven_multilingual_v2 for stable, production-ready output in 29 languages. Choose eleven_flash_v2_5 or eleven_turbo_v2_5 when you need SSML break and phoneme tags or low latency, and eleven_v3 only when maximum expressiveness justifies extra retakes.

With flash or turbo models, insert SSML break tags like <break time="1.5s" /> up to 3 seconds. With multilingual_v2 or v3, use blank lines between paragraphs or insert silence in post-processing with ffmpeg, since ellipsis is not a reliable pause.

Yes, use client.voices.ivc.create with one or more MP3 samples for instant voice cloning. Professional Voice Cloning requires a Creator plan or higher, at least 30 minutes of clean audio, and a verification recording before training.

Sound effects are capped at 22 seconds per generation, and music ranges from 10 seconds to 5 minutes. Concurrent request limits depend on your tier, from 2 concurrent TTS calls on Free up to 15 on Scale.

Generate per-scene MP3s and a manifest.json containing each file's duration, then convert durations to frames and use them as Series.Sequence lengths. Add Audio components per scene or a single continuous track, adjusting demo video playbackRate to fit the voiceover duration.