Text to Speech with Multiple Voices

What multi-voice text to speech means here

In a typical text-to-speech tool you paste text, choose one voice and press play. If the text is a conversation, you hear one person reading every part. The usual workaround is to generate each character's lines separately, download a pile of clips and line them up in an audio editor.

AudioDrama works from a script instead. You label each line with the speaker, the model reads the whole scene in a single pass, and the characters take turns in the order you wrote them. The result is one MP3, not a folder of clips. The underlying model is Fish Audio's Drama 3 text-to-speech model, which we call through its API; AudioDrama is an independent product and not affiliated with Fish Audio.

Because the voices are assigned per character rather than per clip, you can change a whole performance by changing one dropdown and generating again. The text stays as it is.

How to give each character a voice

The script format is deliberately plain. Each line starts with the speaker's name and a colon. A line with no name is read by the Narrator.

1. Write the script. One turn per line, in the form Name: line. Names can be up to five words.
2. Let the cast be detected. The generator lists every name in the order it first appears. The Narrator, if used, gets the deep narrator voice by default and the other characters get distinct voices from the list.
3. Change any voice you do not like. Each character has a dropdown with all 16 voices.
4. Generate. The finished MP3 is saved to your private library so you can replay or download it later.

Here is a four-voice example: three characters and a Narrator. It parses into 4 speakers and 7 turns.

Narrator: Aboard the research station Kepler Nine, the night shift had been quiet for six hours.
Ines [tired, half asleep]: Status report. Please tell me the coffee machine is still working.
Okoye [dry, deadpan]: The coffee machine is fine. The airlock is not.
Ines [suddenly awake]: Say that again.
Pell [nervous, speaking fast]: It cycled itself. Twice. I didn't touch it, I swear I didn't touch it.
Narrator: Nobody spoke. Somewhere below them, a pump started to hum.
Okoye [slow, deadly serious]: Then someone else is on this station.

Names are matched exactly, including capital letters, so Okoye and OKOYE would become two different characters with two different voices. Pick one spelling and keep it. Our guide to the audio drama script format lists the other small rules the parser follows.

The 16 voices

These are the built-in voices, with the descriptions used in the app. They are English voices, and there is no voice cloning.

To hear several voices in one take, play the four-voice radio play sample or the lighthouse scene. To hear a specific voice, sign in and generate a short test: the 2 free minutes are enough for a few lines.

Telling characters apart by ear

Listeners tell voices apart by pitch, pace and accent, not by name, so cast against type contrast. Give two characters who share a scene a clear difference: a deep voice next to a young one, a British accent next to a US one. If you have a narrator plus several characters, keep the narrator the most neutral voice in the cast so the characters stand out against it. Avoid giving similar voices to characters who answer each other directly, because the listener will lose track of who is talking.

Limits you should know about

Multi-voice text to speech is only useful if you know where it stops. These are the actual limits today:

• Script size: up to 6,000 characters per script, and a per-job cap of 9,000 bytes once the speaker tags that the system adds are included. English usually hits the character limit first; Chinese, Japanese and emoji use 2–4 bytes per character and hit the byte cap sooner. Split a long story into scenes or chapters.
• Cast size: at most 16 characters per script, counting the Narrator.
• One speaker per line: there is no way to change voice halfway through a line, and no overlapping speech.
• No sound effects and no music. AudioDrama produces the voice track only. Add effects in an audio editor.
• Free minutes: 2 minutes of audio once, when you sign in with Google. After that you buy credits.
• Credits: counted by audio length or script length, whichever is greater, at 1 second per 15 bytes of script text including directions. This script is 577 bytes once speaker tags are added, so it needs at least 39 seconds of credit. Failed generations are refunded in full. See pricing for the details.

Where multiple voices help

• Story drafts and bedtime stories: let the Narrator carry the prose and give each character a voice, then listen before you commit to a recording.
• Listening practice: write a conversation between two or three speakers for learners. The voices are English, so other languages may work but are not guaranteed.
• Game and animation prototypes: placeholder dialogue for scenes before real actors are booked.
• Explainers with a host and an expert: see the page on dialogue text to speech for two-speaker scripts.

To shape how each character sounds, add a delivery note in square brackets. The page on text to speech with emotions shows which directions to try.