Multi-Speaker Text to Speech: Create AI Conversations
Multi-speaker text to speech turns a script with several voices into one conversation. See how it works, compare limits as of October 2026 and write natural dialogue.
Multi-speaker text to speech is AI voice generation for a script with two or more speakers. Each line belongs to one speaker, and each speaker has its own voice. Limits on speaker count and text length differ from tool to tool, so check them before you write a long scene.
How does multi-speaker text to speech work?
Most tools represent a conversation as a list of turns. Each turn has a speaker and some text, and the speaker points to a voice. Google's speech generation documentation, updated 2026-10-12, describes the same pattern. Every turn must include a speaker that matches a configured speaker, and the single-request mode "supports up to 2 speakers using prebuilt voices" (Google's speech generation docs).
There are two ways to build a scene:
- One request: the tool generates every turn together and returns one audio file.
- Stitched generation: you generate turns separately and join the audio. Google's documentation recommends this approach for custom voices in a multi-character dialogue, where each turn is synthesized on its own and the audio is concatenated.
OndaLabs uses the two-speaker model. Each of the two speakers has a name and a voice, and every line belongs to one of them. Your cloned and designed voices can be used as speakers too, which helps when you keep returning to the same pair of characters.
Multi-speaker text to speech tools compared
The table below lists published limits as of October 2026. Check the vendor page before you plan a project, because limits change.
| Tool or model | Speakers per request | Input rule (as published) | |---|---|---| | OndaLabs Dialogue | Two | 5,000 characters per request; speaker names use letters and numbers, up to 30 characters | | Google Gemini speech generation | Up to two, with prebuilt voices | For custom voices, synthesize each turn and join the audio | | Google Cloud Gemini-TTS | Examples use two; a maximum is not stated | Dialogue field up to 4,000 bytes; output audio about 655 seconds at most | | ElevenLabs Text to Dialogue | No stated limit on speakers | About 2,000 characters total per request, so split longer scripts | | Microsoft Azure Speech (SSML) | Several voices in one document | Markup-based, with voices set inside one SSML document | | VibeVoice-TTS-1.5B (open source) | Up to four | Up to 90 minutes; research and development use only |
Sources: Google Cloud's Gemini-TTS docs, ElevenLabs Text to Dialogue, Microsoft Learn on speech synthesis markup, and the VibeVoice repository. The VibeVoice row needs a caution. Its README says the TTS code was removed on 2025-09-05, the model is intended for research and development only, and it is not recommended for commercial use without further testing.
How to write a dialogue script that sounds like two people
Good dialogue scripts are easier to direct when the structure is simple. These habits help most:
- Give each speaker a short, distinct name. In OndaLabs, names use letters and numbers only, up to 30 characters, and the two names must differ.
- Choose voices with clearly different character. ElevenLabs advises assigning distinct voices to each speaker. OndaLabs offers 30 prebuilt voices, so you can pick two that contrast, such as Kore (firm) and Puck (upbeat).
- Keep turns short. A short turn is easier to direct, and one emotion per line usually reads more naturally than a paragraph.
- Keep each direction short. The product tip reads "Keep directions short: one or two words per line."
- Spell out delivery where it matters. ElevenLabs' best practices say explicit cues give more predictable results than relying on context alone.
Make the conversation feel real: reactions, pauses and laughs
Real conversations contain small moments, such as a quick agreement or a pause before an answer. Google's documentation handles these with pipes for backchannels, for example |oh really?|, and with tags for pauses and sounds. OndaLabs works the same way for reactions. You put the other speaker's reaction inside a line between pipes. The product's own example is "The dragon is non-negotiable. |hmm| Honestly though?"
For timing and sound, use the tags from the 33-tag set. A <short pause> lets a line breathe, a <long pause> marks a real change of topic, and a <laugh> or <chuckle> gives a reply some warmth. Use them sparingly. A scene with a tag in every line sounds like a performance, not a conversation.
Reactions are short moments inside a speaker's line. This post does not claim that OndaLabs mixes overlapping speech. If you need true overlaps, plan to edit the audio in a separate tool.
Limits to plan around
Three limits shape most projects:
- Five thousand characters per request. The counter adds the speaker's name length, plus 2, plus the text length, for every turn. So a turn by "Maya" with 40 characters of text counts as 46 characters.
- Credits follow characters. Onda Lite costs 1 credit per character and Onda Flash costs 1.5, rounded up. The Maya turn above costs 46 credits on Onda Lite or 69 on Onda Flash.
- Long scenes need chunks. Google Cloud's documentation caps output audio at about 655 seconds per request and truncates longer input. Split a long scene into scenes, generate each one, and join them in your editor.
Where AI conversations are used
Two-voice audio fits a wide range of work: explainers hosted by two people, interview practice, tutorials with a narrator and a second voice, game scenes, and language practice with a conversation partner. In each case, short scenes are easier to check line by line than one long file. For game work, see AI voices for games. For recurring characters, see how to create AI character voices for stories.
Try a two-speaker dialogue in OndaLabs
The Dialogue studio gives each speaker a name, a voice and a style direction, and it lets you add tags to any line. The free plan includes dialogue on Onda Lite with 2,000 credits a month, which is about 2 minutes of audio. Onda Flash, the most expressive model, is available on paid plans. You can compare the plans on the pricing page, or sign up and try a short scene first.
Frequently asked questions
How many speakers can an AI text to speech tool handle?
It depends on the tool. Google's Gemini speech generation supports up to two speakers with prebuilt voices in one request. ElevenLabs' Text to Dialogue states no limit on speakers but caps each request at about 2,000 characters. OndaLabs Dialogue uses two speakers and accepts up to 5,000 characters per request.
Can I use my own cloned voice in a multi-speaker conversation?
Yes, in OndaLabs. The Dialogue studio says cloned and designed voices work there too. Clone only from a recording you have permission to use. Custom voice slots start on the Creator plan, with 3 slots, and rise to 10 on Pro and 30 on Business.
Is there a free multi-speaker text to speech tool?
OndaLabs has a free plan that includes dialogue on Onda Lite, with 2,000 credits a month and about 2 minutes of audio. Each dialogue request is capped at 5,000 characters. This post did not verify free tiers from other vendors, so check their current terms directly.
Can AI voices react to each other?
Partly. In OndaLabs, you can place a short reaction inside a speaker's line, such as |oh hmm|. Google's documentation uses the same pipe syntax for backchannels. For true overlapping speech, edit the audio in a separate tool.
How do I write a dialogue script for AI voices?
Give each speaker a short name and a distinct voice, keep turns short, and add one short direction per line. Use reactions and pauses sparingly. Generate a few lines first to check the voices, because every request is charged by characters.