A charcoal studio microphone beside ivory and copper sound-wave ribbons, with one copper segment set apart.
Changing one line is a useful test of how much control a voice tool gives you.Sint

Breakdown / AI audio · Voice generation

Making stories with AI voices: the second take matters

ElevenLabs and Google let you direct a performance. Suno adds music. The useful test is what happens when you change one line.

“I'm fine” can mean several things. Someone might say it warmly to reassure a friend, mutter it through clenched teeth or rush through it because they're clearly not fine. The words stay the same. The performance tells the story.

AI voice tools let you supply both: the script and instructions for how it should sound. For a short film, game or audio story, that opens up a practical creative process. Write a scene, choose the voices and try different deliveries. Then comes the harder part: changing one line without losing everything you liked about the first take.

Give the voice a part to play

ElevenLabs released Eleven v4 on September 28, alongside v4 Turbo for interactive applications. The company positions v4 for dialogue and long-form audio. Its tools can use preceding or following text and references to earlier generations to help sections sound continuous. That's a useful feature when you're assembling a scene from separate takes, though the result still needs a careful listen. ElevenLabs release notes

Google's Gemini 3.8 Flash TTS, announced September 23, supports voice design, two-speaker scenes and directions for individual lines. TTS means text-to-speech: written words go in, spoken audio comes out. Google also offers Flash-Lite TTS, which supports performance direction and is aimed at higher-volume uses. These are the company's stated capabilities, not our comparison of their sound. Google's announcement

For our “I'm fine” example, you might ask for a character who is reassuring someone while hiding their own fear. That gives the model more context than “sound sad.” It's an illustrative direction, not a take we've generated. The question is whether you can hear the difference without the performance becoming exaggerated.

Put music under the words

Suno Speech combines spoken audio with original background music. You provide writing or an idea, then describe the voice and musical style. Suno opened the beta to everyone on October 1, while acknowledging that accents can drift and dramatic pauses can become excessive. Suno's announcement

Picture a message from a missing spaceship: a composed voice reading a routine report while the music slowly becomes unsettling. Or an introduction to a tabletop game, with a narrator setting the scene over a quiet score. Those are briefs you could try, rather than results we're claiming to have heard.

Combining voice and music also creates an editing question. What happens when you love the voice and want a different score? Before building a project around a tool, check what you can change separately.

There are local options, too

Qwen3-TTS offers downloadable models for voice design, preset voices and reference recordings. These are separate variants. The VoiceDesign model's card lists an Apache 2.0 license and explains how to describe a voice in ordinary language. Qwen model card

Kyutai's Pocket TTS is designed to run on a computer's CPU. Its repository now includes smaller English models and standalone downloads as well as the Python version. The individual voices have their own licensing information, so the model's availability doesn't answer every question about the voice you choose. Pocket TTS repository · Voice terms

Start with a scene you can finish

Try a short passage with a clear change in mood. Listen for missing words, odd emphasis and a voice that changes halfway through. Then replace one sentence and see how much work it takes to make the new take fit. Count the attempts as well as the price of each generation.

We haven't run a head-to-head listening test of these models. This is a guide to their documented controls and the questions to ask of them. The most revealing test may be the second take: ask for a small change, then listen to what else changed with it.

Sources