Adding a voiceover with ElevenLabs

A voiceover is speech recorded apart from the picture and laid over it. In a generated film it carries the words, while the shots carry the place and the mood, and keeping the two apart means you can change a line without rendering the clip again. ElevenLabs is a company that makes speech models. You type a line, choose a voice and get back an audio file. There are two ways in. The web app lets you paste text, press generate and download the result, and the API does the same from a script, which is how you make thirty lines at the same settings and keep a record of every one.

Write the script for the ear and for the cut. Each line should belong to one shot, so an 8-second Veo clip gets one short sentence with room around it, and the voice should start about half a second into the shot and finish before the cut. Say it aloud with a stopwatch before you generate anything. Spoken sentences run shorter than written ones, and a line that needs a comma to make sense on paper is often better as two lines. A 24-second product film might carry three lines, such as 'Made in one small mill.', 'Washed twelve times before it reaches you.' and 'Forty pounds, and it lasts.' Each one becomes its own audio file, named after the shot it sits under.

ElevenLabs' model page sorts its text-to-speech models by what each is for, and two of them suit a voiceover. eleven_v3 is the expressive one, reads more than 70 languages and takes up to 5,000 characters a request, while eleven_multilingual_v2 covers 29 languages, takes 10,000 characters and is the one the docs call most stable on long-form work. The Flash models are built for speed. Name the model every time, because the API reference defaults to eleven_multilingual_v2 when you leave it out and the quickstart's own sample asks for eleven_v3.

A request sends the text, a voice ID, a model ID and any voice settings, and the audio comes back in pieces that your code writes to a file as they arrive. The key travels in a header called xi-api-key, and the Python SDK reads it from ELEVENLABS_API_KEY. Five settings shape the read. Stability starts at 0.5, similarity_boost at 0.75, style at 0 and speed at 1.0, which the best-practice guide lets you move between 0.7 and 1.2, and use_speaker_boost is on. Eleven v3 behaves differently, since the app offers its stability as three named positions rather than a number, with Creative the loosest and Natural the closest to the original recording, and it takes stage directions in square brackets such as [whispers] or [sighs].

The API bills by the character. ElevenLabs' API pricing page puts eleven_v3 and Multilingual v2 at $0.10 per 1,000 characters and Flash at $0.05, so three short lines of voiceover, 100 characters in all, cost one cent. Plans start at $6 a month for Starter, then $22 for Creator (the first month is $11) and $99 for Pro. The app counts credits instead, one per character on Multilingual v2, and its pricing page gives Starter 30,000 credits where the API page gives the same plan 60,000 characters of v3, so check which meter a plan feeds before you budget. Free accounts may only use the audio for non-commercial work.

Check every take before it goes under a picture. Listen once on headphones and once on a phone speaker, and listen hardest for names, numbers and the final word. When a name comes out wrong, ElevenLabs' own guide says to try spelling it the way it sounds, and for the API it offers pronunciation dictionaries with alias tags, where its example tells the model to read Claughton as Cloffton. Numbers need the same care. The guide notes that Multilingual v2 reads $1,000,000 as one million dollars while Flash v2.5, a much smaller model, reads it as one thousand thousand dollars, which is why the scripts in this module spell prices out.