Write the script, cast it with text-to-speech voices, then add a score and sound effects to match each scene. Everything runs in one workspace, so you can hear a rough episode the same afternoon and fix only the lines that sound wrong.
An audio drama is a film with the picture removed, so every job the camera used to do now belongs to a sound. Voices carry the characters, music carries the mood, and a door slam has to tell the listener they are standing in a hallway. AI can now supply all three layers from a script you write: a cast of voices, a score and spot effects. This page covers what to use for each layer, how to put them together, and where you will still need an audio editor.
Audio fiction has always had a staffing problem. A scene with four characters, a thunderstorm and a creaking staircase traditionally needs actors, a studio, a foley table and someone to mix it all. That is why most people who write dialogue well never get to hear it performed. Text-to-speech models remove the first hurdle: you paste a line, pick a voice and get a performance back, and you can repeat it as often as the script needs without scheduling anyone.
Picasso IA puts the pieces in one place. The catalog holds 502 models, including 24 text-to-speech models and 10 AI music models, and the toolkit is the workspace where you pick one, write a prompt or attach a file and generate. New accounts get free credits, so a test scene costs you nothing to start, and the pricing page explains how credits are spent once those run out.
The honest advantage is iteration speed on the writing side. A script that reads well on the page often falls apart when spoken: a line is too long, two characters sound identical, a joke needs a beat of silence. Hearing it exposes all of that within minutes.
Dialogue is the layer listeners judge hardest, so start there. Two models in the catalog suit drama well, and they behave differently enough that it is worth testing both on the same scene.
elevenlabs/v3 offers a long list of named voices, speed control, a style setting that moves delivery from flat narration toward something more theatrical, and a stability setting meant to keep a voice behaving the same across a long script. It also accepts the text before and after the current line, which helps the intonation at the start and end of a sentence make sense in context. inworld/tts-1.5-max works through markup inside the script itself: tags such as [happy] or [sad] steer the emotion of a line, [laugh] and [sigh] add non-verbal sounds, and break tags place pauses exactly where you want them. It also lets you choose preset voices by name.
Whichever you use, assign each character one voice and write the choice down. A cast sheet with the character, the model, the voice name and the settings is the single most useful document in the whole project, because you will regenerate lines later and need to match what you made on day one. Give a narrator a voice clearly different from every character.
If you want a character to use your own voice, the voice cloning page covers that, and it applies only to a voice you own or have permission to use. Fictional characters get fictional voices, never an imitation of a real, identifiable person.
Music and effects are separate jobs, and the catalog treats them separately. For the score, music models such as elevenlabs/music, stability-ai/stable-audio-2.5 and google/lyria-3 turn a written description into a track. elevenlabs/music has an instrumental switch and a target length, which is exactly what you want for a bed that sits under dialogue. Ask for a mood, a few instruments and a length, and keep vocals off so the music never competes with a character speaking.
Sound effects are less direct. The catalog has no model built only for text-to-sound-effect, so you have two routes. stability-ai/stable-audio-2.5 lists short one-shot effects among its outputs, so describing a single event, such as a heavy wooden door closing, is worth trying. The other route is picture-led: mirelo/video-to-sfx-v1.5 and zsxkib/mmaudio read a video clip and generate sound that matches what is on screen. That suits you only if the scene already exists as video, but it is handy if you are also making a visual version. For more on that route, see adding sound effects to silent AI videos.
| Layer | Models to try | What you give it | What to watch for |
|---|---|---|---|
| Dialogue | elevenlabs/v3, inworld/tts-1.5-max | One line, a voice, a delivery note | The voice can drift between runs |
| Narration | elevenlabs/v3, inworld/tts-1.5-max | A short passage in a distinct voice | Long passages may need splitting |
| Score | elevenlabs/music, google/lyria-3 | Mood, instruments, length, instrumental on | Music does not know where your lines fall |
| Spot effects | stability-ai/stable-audio-2.5 | One sound event per prompt | Results vary, so generate a few |
| Synced effects | mirelo/video-to-sfx-v1.5, zsxkib/mmaudio | A video clip of the scene | Needs picture, and works on short clips |
Plan on a short first episode. Two or three characters and one location is enough to learn the whole pipeline before you scale up.
The most common mistake is generating a whole scene in one pass. A single long block of text gives you one delivery and no control over who sounds how. Generate each line separately, so you can redo one bad take without touching the rest, and so the pauses between speakers are yours to set rather than the model's.
The second is writing for the eye. Models read punctuation as performance, so a wall of commas produces breathless speech while short sentences give the voice room to land. Read your script aloud once before generating anything, and cut every line you stumble over. A character also cannot perform an action that exists only in a bracket, so turn it into a sound or a line.
The third is burying the dialogue. Music generated at full energy will drown a quiet voice, and it will do so in the mix, where fixing it is easy only if the tracks are separate. Keep each layer in its own file, and ask for instrumental music with a calm arrangement.
Tip: keep one sound per effect prompt and one voice per file. A single clean door slam you can place anywhere is worth more than one long generation of door, footsteps and wind that you cannot separate afterwards.
This workflow gives you a convincing rough cut, and sometimes a releasable episode, but it has real limits that are worth knowing before you commit weeks to a series.
The AI handles the performance and the raw material; you handle the structure, the timing and the final polish.
It can generate the voices, the music and the effects, but it does not assemble them for you. You write the script, generate each line, add a score and effects, and then arrange the files on tracks in an audio editor. Picasso IA is the cast and the sound library; the editor is the mixing desk. A short episode with two or three characters is a realistic first goal.
Pick one voice for each character and never change it midway. Keep the settings fixed too: with elevenlabs/v3 that includes the stability and similarity controls, and with inworld/tts-1.5-max it means the same preset voice and speaking rate. Generate lines in the same sitting when you can. If one line still sounds off, regenerate that line alone instead of redoing the scene.
Neither is best everywhere, so test both on your hardest scene. elevenlabs/v3 has many named voices and a style control for more theatrical delivery. inworld/tts-1.5-max lets you write emotion and non-verbal sounds straight into the script with tags. Choose by ear for your characters, and use one model per character to avoid mismatches. The catalog has 24 text-to-speech models if neither fits.
There is no model in the catalog built only for that, but stability-ai/stable-audio-2.5 lists short one-shot effects among its outputs, so describing one sound event per prompt is worth trying. Results vary, so generate several versions and keep the best. If you have the scene as video, mirelo/video-to-sfx-v1.5 and zsxkib/mmaudio generate effects that match the picture instead.
Yes. Music models such as elevenlabs/music, stability-ai/stable-audio-2.5 and google/lyria-3 create tracks from a description. For drama, ask for an instrumental with a calm arrangement and a length that covers the scene, then lower its volume under the voices when you mix. Generate a few options per scene.
Yes, for any episode with more than one layer. Picasso IA generates the voices, music and effects as separate files, but it does not arrange them on a timeline. A free editor is enough for trimming, placing tracks, setting volumes, fading music in and out and exporting the final file. The skill you need is listening, more than technical knowledge.
You can, if the voice is yours or you have clear permission from its owner. Voice cloning is covered on the AI voice cloning page. Never clone or imitate a real person without their consent, and keep cloned voices for characters you control. Fictional characters can simply use the preset voices the models offer.
Short enough to finish. A few minutes with two or three characters and a single location lets you test the voices, the score and the mixing without the project growing out of control. Once it sounds right, you have a repeatable process and a cast sheet.
Often yes, but check the current requirements of the platform you publish on, since podcast hosts, video sites and streaming services each have their own rules about AI-generated audio and disclosure. Also check the terms of each model you used. Keep your characters fictional, and credit the production honestly if the platform asks you to.
It depends on the models and settings you choose and on how many takes you generate, since every generation uses credits. Dialogue is the biggest count, since you generate line by line. New accounts start with free credits, and the pricing page is the only place to check current plans.
Ready to hear your script out loud? Open the Picasso IA toolkit, pick a voice and generate the first line of your first scene.
Choose your language
