How to Make a Two-Voice AI Podcast From a Script
A dialogue-capable text-to-speech model can turn a script labeled "Host:" and "Guest:" into one finished two-voice audio file in a single generation — no recording, no stitching two clips together, and no manual timing.
You can turn a plain-text script into a finished two-voice conversation in one generation, with no recording booth and no manually stitching two single-voice clips together: label each line with a speaker's name (Host:, Guest:), assign each name a distinct voice, and a dialogue-capable text-to-speech model renders the whole exchange — pacing, turn-taking and all — as a single audio file. Google's Gemini Flash TTS, available through Arcframe, does exactly this for up to two speakers and 87 languages.
What "two-voice AI dialogue" actually means
Most text-to-speech tools generate one voice per request. If you wanted a two-person podcast intro, the usual workaround was generating each speaker's lines separately and editing them together by hand — lining up pauses, fixing overlaps, exporting twice. A dialogue-capable model skips that: it reads a single script where each line is prefixed with who's speaking, and produces one audio file where both voices already sound like they're in the same room.
The script format the model expects
The format is just plain text with a speaker label at the start of each line:
Host: Welcome back to the show. Today we're talking about how AI can generate a two-person conversation from nothing but a script.
Guest: Thanks for having me — this is one of those things that sounds like a gimmick until you hear it actually hold a rhythm.
Each speaker name is then mapped to a voice — for example {"name": "Host", "voice": "Charon"} and {"name": "Guest", "voice": "Kore"}. The model supports exactly two speakers per generation, not three or more. You can also drop in delivery cues like [sigh], [whispering] or [short pause] inline, and a free-text style instruction (up to 200 characters) to steer overall tone — "a relaxed, slightly informal tech podcast" reads differently than "a dramatic newscast."
Three ways to build a two-voice episode, compared
| Method | Setup | Turn-taking | Editing needed | What drives cost |
|---|---|---|---|---|
| Two single-voice clips stitched together | Two separate generations | Manually timed by you | High — aligning pauses, avoiding dead air or overlap | Character cost of both clips, plus your editing time |
| One script in a native dialogue model | One generation | Handled automatically from the speaker labels | None | Flat rate per 1,000 characters of the whole script |
| Record two real people | Full production — mics, room, schedules | Natural | Audio cleanup and mixing | Studio time and editing labor |
Picking two voices that don't sound alike
Because there's no visual cue for who's talking, contrast matters more than either voice being "good" on its own. A firm voice against a warm one reads more clearly than two similarly bright voices. A few of the presets available:
| Voice | Character |
|---|---|
| Charon | Informative |
| Kore | Firm |
| Sulafat | Warm |
| Puck | Upbeat |
| Fenrir | Excitable |
| Umbriel | Easy-going |
| Algenib | Gravelly |
| Zephyr | Bright |
These are preset voices, not clones of anyone's actual voice — which is also a limitation worth naming (see below).
Language support
Gemini Flash TTS covers 87 languages in total. Thirteen of them are tagged specifically to India — Hindi, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, Konkani, Maithili, Sindhi and English (India) — and Bengali and Urdu are supported as well, listed under Bangladesh and Pakistan respectively. A Hindi or Tamil two-voice dialogue works the same way as an English one: label the lines, pick voices, generate. Machine-translating a script line by line before generating tends to sound stiff, though — a human pass over the translation to keep it conversational is worth the extra step.
Where this breaks down
- Exactly two speakers, not three or more. A roundtable with three hosts doesn't fit this mode.
- No cloned voices in dialogue mode. Both speakers are chosen from the preset voice list above; you cannot drop in a cloned voice of yourself or a colleague the way you can for single-voice narration.
- Audio only. This produces a sound file, not a video. Pairing it with visuals is a separate generation step.
- Cost tracks script length, not runtime. Pricing is per 1,000 characters of the input script, not per minute of output audio, so a long, talkative script costs more regardless of how fast the voices end up speaking it.
Step by step
- Write the script as plain lines, each starting with the speaker's name exactly as you'll assign it —
Host:,Guest:. - Pick two contrasting voices from the preset list so listeners can tell speakers apart by ear alone.
- Optionally add a style instruction for overall tone, and inline cues like
[short pause]where you want a beat. - Generate. One audio file comes back with both voices in the finished conversation, already timed to the script.
- For another language, translate the script (with a human tone pass) and generate again — the same two-speaker format applies.
This is one of the audio models available inside Arcframe alongside video, image and 3D generation, so a two-voice dialogue can sit in the same project as the video or image work around it. The free plan includes 20 one-time credits with no card required, and generation costs for every model are listed on the pricing page.