AI Voice for Podcasts
Published September 2, 2026 · Every claim below was read from the vendor’s own documentation on September 2, 2026. This page is about whether a provider can produce a conversation at all, not a ranking — for that, see the flagship comparison. No listening test has been run.
The short answer
Most podcasts have more than one voice, and most of these providers cannot produce more than one voice in a request. Only three of the ten document multi-speaker output, and that is the constraint that decides this use case before any other.
ElevenLabs publishes a dedicated dialogue endpoint taking a list of turns, each with its own text and voice, capped at “10” unique voice IDs.
Google supports it only on one model family. Multi-speaker is documented for Gemini-TTS models, using a multiSpeakerMarkup structure of named turns — and not for the rest of its catalog.
Azure does it through plain SSML, with multiple voice elements inside one speak block, and adds something nobody else offers: “These voices can be in different languages.”
For the other seven, a two-host episode means one request per turn and stitching the audio yourself. That is workable. It is just not what the marketing implies.
Basis: vendor documentation only, checked September 2, 2026 (how we verify).
On this page
Who can produce a conversation
| Provider | Multi-speaker in one request? | Mechanism |
|---|---|---|
| Amazon Polly | Not found | SSML supported, but no multi-voice mechanism documented |
| Azure Speech | Yes | Multiple voice elements inside one speak; voices may be in different languages |
| Cartesia | Not found | One voice per request; SSML-like tags control delivery, not speakers |
| ElevenLabs | Yes | Text to Dialogue: a list of turns, each with text and a voice ID; maximum 10 unique voices |
| Google Cloud | Yes, on Gemini-TTS only | multiSpeakerMarkup with named turns |
| MiniMax | No — see below | timbre_weights blends up to four voices into one, which is a different thing |
| Murf AI | Not found | A single voiceId per request |
| OpenAI | Not found | A single voice per request from a built-in list |
| Resemble AI | Not found | Also no longer sells voice generation |
| Speechify | Not found | SSML for delivery; no multi-voice mechanism documented |
How each one does it
ElevenLabs is the only vendor with a purpose-built dialogue surface. Its parameter is documented as: “inputs (list of object, required) — A list of dialogue inputs, each containing text and a voice ID which will be converted into speech. The maximum number of unique voice IDs is 10.” Delivery is controlled per turn: “In Text to Dialogue, each dialogue turn has its own text and voice. Add audio tags inside the text for the turn they should affect. The voice_id still selects the speaker voice for that turn, while the tags guide delivery.” For a scripted two-host show, that is the shape of the problem exactly.
Google’s support is real but narrowly scoped. Its documentation lists “Speaker number support Single, multi-speaker (dialogue)” gated to the Gemini-TTS models, with a markup structure of named turns: "multiSpeakerMarkup": { "turns": [ { "speaker": "Sam", "text": "Hi Bob, how are you?" }, { "speaker": "Bob", "text": "I am doing well, and you?" } ] }. If you had picked a Google voice from another family for other reasons, multi-speaker is not available to you on it.
Azure needs no special endpoint. Verbatim: “Within the speak element, you can specify multiple voices for text to speech output. These voices can be in different languages. For each voice, the text must be wrapped in a voice element.” Its worked example puts two named voices in one block. The cross-language clause is genuinely distinctive: a bilingual interview segment is a documented capability rather than a workaround.
One thing that looks like it and is not
MiniMax publishes a parameter that reads like multi-speaker support and is not. Verbatim: “The weight assigned to each voice. Must be specified together with voice_id. Supported range: [1, 100]. Up to 4 voices can be mixed. A higher weight value increases similarity to the corresponding voice.”
That is voice blending — combining several voices into one new voice — not two people talking. It is a useful feature for designing a host voice. It will not give you an interview.
If your provider does one voice at a time
Seven of the ten leave you to assemble the conversation. That is entirely workable and it is how most AI-voiced shows are actually made, but it changes your production in three ways worth planning for.
- Turn boundaries become your responsibility. Nothing in the audio knows a reply follows a question, so pacing between turns is something you set in the edit rather than something the model produces.
- Consistency across many small requests matters more. Each turn is a separate generation, and no vendor guarantees that identical input yields identical output — none documents a seed. Whether turn 40 sits comfortably beside turn 1 is worth testing before you script a series.
- Your request count multiplies. A 40-minute two-hander is hundreds of requests, which puts you against concurrency limits documented in our API comparison — where free tiers run from 1 to 5 simultaneous requests.
Episode length against request limits
An episode script of 40 minutes is roughly 6,000 words, or about 36,000 characters. Almost nobody takes that in one request: OpenAI caps at 4,096 characters, Amazon at 6,000, Speechify at 2,000 on its standard endpoint and 20,000 on its streaming one, and ElevenLabs runs 5,000 to 40,000 depending on model.
For a dialogue show this matters less than it does for audiobooks, because you are already generating turn by turn and turns are short. It matters most for monologue segments — a ten-minute solo intro is one long block of text, and that is where the cap bites.
What you must tell listeners
Three vendors place a positive duty on you to disclose that the voice is synthetic: OpenAI requires “a clear disclosure to end users that the TTS voice they are hearing is AI-generated and not a human voice”; Microsoft requires disclosure of “the synthetic nature of generated voices”; MiniMax requires a prominent mark within the content. Several vendors also embed watermarks in what they generate. The obligations are set out in full on our page on publishing with AI voice, and they apply to a podcast as much as to a video.
Frequently asked questions
Can I make a two-host podcast with AI voices?
Yes, but only three providers do it in a single request — ElevenLabs through a dedicated dialogue endpoint, Google on its Gemini-TTS models only, and Azure through multiple voice elements in SSML. With the rest you generate each turn separately and assemble the episode yourself.
How many speakers can I have?
ElevenLabs documents “The maximum number of unique voice IDs is 10” for its dialogue endpoint. Google and Azure publish no explicit speaker cap in what we read.
Can two speakers be in different languages?
Azure documents it: “These voices can be in different languages.” We found no equivalent statement from the other two multi-speaker providers.
Is MiniMax’s voice mixing the same as multi-speaker?
No. timbre_weights blends up to four voices into a single new voice. It is voice design, not conversation.
Do I have to tell listeners the hosts are AI?
Your vendor may require it. Three of the ten place a positive disclosure duty on you. Several also watermark the audio, which makes the synthetic origin detectable regardless.
Which sounds most natural in conversation?
We have not tested that and do not rank it. Vendors make claims — ElevenLabs markets “Support for natural multi-speaker dialogue” — and what each charges for such claims is documented separately.
Sources and what we could not verify
Every claim was read from the vendor’s own documentation on September 2, 2026. Request limits and concurrency are set out with sources in our API comparison; the method is on the methodology page.
What we could not verify
- Whether multi-speaker output sounds like a conversation. No listening test has been run. We report that the mechanism exists, not that it works well.
- Speaker limits for Google and Azure. Neither publishes a maximum we could find.
- Whether separately generated turns sit consistently together. No vendor documents consistency between requests, and none offers a seed.
- Whether the seven providers without a documented mechanism support it undocumented. Absence in the documentation we read is not proof of absence.
Change log
September 2, 2026 — First publication. All claims read from official vendor documentation on September 2, 2026 and dated accordingly.
Published September 2, 2026 · This page shows no audio and reports no listening results, because our benchmark has not run. Corrections are recorded with their date on the corrections page. Independence note: this page contains no affiliate links, and no vendor paid for placement or influenced the order — see how we make money and our editorial policy.