Text to Dialogue is a dedicated ElevenLabs API endpoint that generates multi-speaker conversational audio from structured text input. Where standard Text to Speech takes a single string of text and generates audio in one voice, Text to Dialogue takes an array of text-voice pairs — each element specifying what is said and which voice says it — and generates a single cohesive audio file containing all speakers in natural conversation.
The feature is available in the ElevenLabs web dashboard as Dialogue Mode (activated when multiple speakers are added to a TTS session) and via the API at the POST /v1/text-to-dialogue/convert endpoint. It requires the Eleven v3 model — the only model that supports multi-speaker dialogue generation with the contextual awareness required for natural conversational prosody.
The feature is still under active development as of May 2026. ElevenLabs notes in documentation that actual results may vary, and certain behaviours — particularly around overlapping speech and interruptions — are nondeterministic and may require the seed parameter for more reproducible output.
How Text to Dialogue Works: The API Structure
Basic Request Format
The Text to Dialogue API accepts a JSON body with a required inputs array. Each element in the inputs array is an object containing text (the speech content for this turn) and voice_id (the ElevenLabs voice ID of the speaker). An optional model_id defaults to eleven_v3. Optional parameters include seed (integer 0–4,294,967,295 for determinism), language_code (ISO 639-1 for language enforcement), and pronunciation_dictionary_locators (up to 3 pronunciation dictionaries for custom word pronunciations).
Example Request Structure
A basic two-speaker dialogue: inputs array containing: {text: ‘So you actually think AI is going to replace voice actors completely?’, voice_id: ‘VOICE_ID_SPEAKER_1’}, {text: ‘[laughs] Replace them? No. Change how they work? Absolutely.’, voice_id: ‘VOICE_ID_SPEAKER_2’}, {text: ‘That sounds like a very diplomatic answer.’, voice_id: ‘VOICE_ID_SPEAKER_1’}, {text: ‘[sighs] It is, because the honest answer is more complicated.’, voice_id: ‘VOICE_ID_SPEAKER_2’}. This single API call generates a complete 4-turn dialogue with both speakers, including the laughs and sighs Audio Tags delivered in context.
Key Technical Constraints
| Parameter | Limit | Note |
| Unique voices per request | Maximum 10 | More than 10 unique voice IDs returns an error |
| Total character count | 2,000 characters recommended | Longer content may truncate or return validation error; split and concatenate |
| Model | Eleven v3 only | Text to Dialogue is not available on Multilingual v2 or Flash models |
| Determinism | Nondeterministic | Use seed parameter for more consistent results across regenerations |
| Audio events | Supported via Audio Tags | Environmental events like [crowd applause] in addition to emotional tags |
Audio Tags and Audio Events in Dialogue
Emotional Audio Tags
Eleven v3’s Audio Tags are fully supported in Text to Dialogue inputs. Each speaker’s text can include directional tags that shape the emotional delivery of their lines: [whispers], [laughs], [sighs], [shouts], [excitedly], [nervously], [sternly]. The tag applies to the text following it in that speaker’s turn, and the model interprets it in the conversational context — a [laughs] tag from Speaker B after an amusing line from Speaker A produces a more contextually appropriate amused delivery than the same tag in isolated TTS.
Environmental Audio Events
Eleven v3 in Text to Dialogue uniquely supports environmental audio events — non-speech sounds inserted into the dialogue that influence the atmosphere of the scene. These are inserted the same way as emotional tags, in square brackets: [leaves rustling], [gentle footsteps], [crowd applause], [football match noise], [typewriter clicking]. Environmental events add acoustic context to dialogue scenes without requiring separate SFX tracks to be layered in post-production. A dialogue scene set in a busy café can include [espresso machine background] and [muffled conversation] events that make the setting audible.
Related: For standalone AI sound effects generation at 48kHz, see our ElevenLabs SFX V2 and SB1 guide
Use Cases
Podcast Production: AI Interview Simulations
Podcast producers using ElevenLabs for content generation can use Text to Dialogue to produce interview-format content between an AI host voice and an AI guest voice — or between a cloned host voice and a designed guest character voice. The conversational naturalness of Text to Dialogue produces interview audio that sounds significantly more like genuine conversation than two separately generated TTS tracks spliced together. For educational podcasts, explainer series, and content marketing audio that uses a Q&A format, Text to Dialogue eliminates the manual synchronisation work of traditional multi-speaker production.
Game Development: NPC Dialogue
Game developers building narrative games with branching NPC dialogue face a specific challenge: generating large volumes of conversational audio between multiple characters quickly and cost-effectively. Text to Dialogue is well-suited for this use case — define character voices for each NPC, write the dialogue exchanges in the API input format, and generate the complete scene audio in a single call. The 10-voice-per-request limit accommodates most NPC conversation scenes. For a game with 50 NPC characters, the per-call limit is not per-project — each dialogue scene can use up to 10 voices, and different scenes can use different voice sets.
Audiobooks with Multiple Characters
Audiobook producers who want distinct voices for each character can use Text to Dialogue to generate character dialogue while using standard TTS for narrator passages. A scene with three characters speaking to each other generates as a single Dialogue API call with three voice IDs — the narrator passages before and after the scene generate as standard TTS. The two audio types concatenate seamlessly. This workflow gives audiobooks character vocal distinction that single-narrator AI audiobooks lack, while keeping narrator consistency through a single TTS voice.
Training and E-Learning Simulations
Corporate training content that uses scenario-based learning — a customer service representative handling a difficult customer, a manager giving difficult feedback, a sales call simulation — benefits specifically from conversational audio rather than narrator-plus-character audio. Text to Dialogue generates the simulation scenario as genuine two-way conversation, making the training content more engaging and more realistic as a practice scenario than a single narrator describing the interaction.
Text to Dialogue vs Standard Text to Speech
| Dimension | Text to Dialogue | Standard Text to Speech |
| Speakers | 2–10 unique voices per request | 1 voice per request |
| Conversational prosody | Yes — speakers influence each other’s delivery | No — each generation is independent |
| Audio Tags | Full support — emotional and environmental | Emotional tags only |
| Output | Single audio file with all speakers | Single audio file with one speaker |
| Model requirement | Eleven v3 only | Flash v2.5, Multilingual v2, Eleven v3 |
| Character limit | 2,000 chars recommended across all inputs | No strict limit — split long content into streaming chunks |
| Best for | Interviews, dialogue scenes, conversations | Narration, monologue, voice agent responses |
| Timeline editing required | No — single API call produces complete scene | Yes — multiple generations require manual sync |
| Determinism | Lower — use seed for consistency | Higher — same text and settings produce similar output |
Prompting Tips for Best Results
Write conversationally, not as formal prose
Text to Dialogue works best with scripts written in natural spoken language — contractions, incomplete sentences, conversational interruptions. Formal or literary prose generates dialogue that sounds read rather than spoken. ‘It is important to note that…’ produces more stilted dialogue than ‘Here’s the thing…’ Write dialogue the way two real people would actually speak, including natural hesitations, responses that reference the previous speaker’s point, and emotional reactions.
Use Audio Tags to cue emotional reactions
The most natural-sounding dialogue uses Audio Tags to mark reactions that real speakers would express — laughter at a joke, a sigh before a difficult admission, an excited uptick at good news. Include these tags in the text of the responding speaker rather than relying on the model to infer the emotional response from context alone. Explicitly tagging reactions produces more consistent and more natural results than leaving emotion inference entirely to the model.
Use the seed parameter for revision workflows
Because Text to Dialogue is nondeterministic, regenerating the same input without a seed produces different audio each time. When you find a generation that is mostly right but has one turn that does not land correctly, you cannot isolate and regenerate just that turn — the entire request regenerates together. Setting a consistent seed for your generation sessions gives you more reproducibility when iterating on dialogue scripts, though ElevenLabs notes that determinism is not guaranteed even with a fixed seed.
Three Insights Most Coverage of Text to Dialogue Misses
1. Conversational Prosody Is the Feature, Not Multi-Speaker Output
Most descriptions of Text to Dialogue focus on the multi-speaker output as the key feature — ‘you can have two voices in one file’. The more important capability is conversational prosody: the way Speaker B’s delivery is contextually shaped by Speaker A’s preceding line. This is what makes Text to Dialogue produce audio that sounds like a conversation rather than two people reading scripts next to each other. It is the same distinction that separates a good actor from a card reader — and it is not achievable by generating two voice tracks independently and placing them sequentially.
2. Environmental Audio Events Are an Underdocumented Capability
Most tutorials and guides covering Text to Dialogue focus on emotional Audio Tags and ignore environmental audio events. The ability to insert [crowd applause], [leaves rustling], [football match noise], and [typewriter clicking] into dialogue scenes is a capability with significant practical value for audio drama, game audio, and immersive content production. A dialogue scene set in a stadium sounds like it is actually in a stadium. A classroom scene sounds like a classroom. This eliminates the need to layer environmental SFX in post-production for many use cases — the atmosphere is built into the dialogue generation itself.
3. The 2,000-Character Limit Is a Design Constraint, Not a Bug
The 2,000-character recommended limit across all inputs in a single Text to Dialogue request appears restrictive for long-form content. It is actually a deliberate design constraint that produces better output — the model maintains conversational coherence more reliably within shorter exchanges than across long scenes. The correct workflow for longer dialogue content is to split the dialogue into natural scene breaks, generate each scene as a separate request, and concatenate the resulting audio files. Each scene maintains full conversational naturalness within its 2,000-character limit. Attempting to push all content into a single long request produces less coherent output than properly segmented generation.
Text to Dialogue in 2027
Text to Dialogue is under active development and its trajectory suggests three expansions. The character limit per request will likely increase as the model’s contextual coherence window expands — the current 2,000-character limit reflects the model’s reliable generation range, and as Eleven v3 successors improve long-context handling, this limit will extend. Environmental audio events will expand from the current set to a broader vocabulary, potentially including spatial audio cues (audio appearing to come from different directions) relevant to immersive audio and game audio production. And streaming output — currently not available for Text to Dialogue (only standard TTS supports streaming) — will eventually extend to dialogue generation, enabling real-time multi-speaker audio for voice agent applications.
Key Takeaways
- Text to Dialogue generates multi-speaker conversational audio in a single API call — up to 10 unique voices, with natural conversational prosody that makes speakers sound like they are genuinely responding to each other.
- Exclusive to Eleven v3 model. Maximum 10 unique voice IDs per request. Keep total character count at or below 2,000 for reliable generation — split longer content into chunks.
- Supports emotional Audio Tags ([laughs], [whispers]) and environmental audio events ([crowd applause], [leaves rustling]) that build atmosphere directly into the generation.
- Use the seed parameter for more reproducible results. Output is nondeterministic — regenerations without seed will vary.
- Best for: podcast interviews, game NPC dialogue, audiobook character scenes, training simulations. Better than manually synchronised independent TTS tracks for any conversational content.
Conclusion
ElevenLabs Text to Dialogue is the most significant addition to the AI audio production toolkit for multi-speaker content creators in 2026. For podcast producers, game developers, audiobook creators, and training content teams, the ability to generate complete conversational scenes in a single API call — with genuine conversational prosody rather than independent voice tracks — removes the most time-consuming manual synchronisation step from multi-speaker audio production. The 2,000-character limit requires thoughtful content segmentation, and the nondeterministic output requires the seed parameter for revision workflows. Within those constraints, Text to Dialogue produces multi-speaker audio that sounds more like genuine conversation than any alternative workflow currently available.
Frequently Asked Questions
What is ElevenLabs Text to Dialogue?
A specialised API endpoint that generates multi-speaker conversational audio from an array of text-voice pairs in a single request. Up to 10 unique voices per request, available exclusively on the Eleven v3 model. Produces more natural-sounding conversation than manually synchronised independent TTS tracks.
How many speakers can Text to Dialogue support?
Up to 10 unique voice IDs per request. There is no limit on the number of turns — the same two voices can alternate across any number of exchanges — but each unique voice used counts toward the 10-voice maximum.
Does Text to Dialogue support Audio Tags?
Yes — full Eleven v3 Audio Tags are supported in each speaker’s text input, including emotional tags ([laughs], [whispers], [excited]) and environmental audio events ([crowd applause], [leaves rustling]).
What is the character limit for Text to Dialogue?
ElevenLabs recommends keeping the total character count across all inputs at or below 2,000 characters per request. Longer requests may truncate or return a validation error. Split longer dialogue content into scene-length chunks and concatenate the audio files.
Is Text to Dialogue available on the free plan?
Text to Dialogue requires the Eleven v3 model, which is only available on paid plans. Free plan users do not have access to Eleven v3 or Text to Dialogue.
Methodology
Text to Dialogue capabilities from ElevenLabs official documentation at elevenlabs.io/docs/overview/capabilities/text-to-dialogue and the Text to Dialogue quickstart. API parameters from ElevenLabs API reference at elevenlabs.io/docs/api-reference/text-to-dialogue/convert. Audio event capabilities from ElevenLabs Eleven v3 documentation. Use case information from ElevenLabs official product pages and editorial team testing. This article was drafted with AI assistance and reviewed by the editorial team at ElevenLabsMagazine.com.
References
ElevenLabs. (2026). Text to Dialogue documentation. https://elevenlabs.io/docs/overview/capabilities/text-to-dialogue
ElevenLabs. (2026). Text to Dialogue API reference. https://elevenlabs.io/docs/api-reference/text-to-dialogue/convert
ElevenLabs. (2026). Text to Dialogue quickstart. https://elevenlabs.io/docs/cookbooks/text-to-dialogue
