Text to Speech with Two Voices: Fish Audio Drama 3 Guide

To make text to speech with two voices, assign one voice to each character and keep that assignment unchanged as you revise. Start with one short exchange: check who speaks each line before directing its emotion.

A believable scene needs more than two different voices. One character can be hiding something while the other tries to sound reassuring. The useful question is how to turn that intention into a script you can generate, listen to and revise.

This guide gives you a small scene, a performance brief, a web-editor workflow, an optional API request and a checklist for taking the dialogue into a video edit.

Pencil0 does not currently integrate Fish Audio Drama 3. Generate Drama 3 speech in Fish Audio. Pencil0 can help with the script, character references and shots, but this article does not describe a built-in Drama 3 button or an external-audio import feature. The examples below are original, documentation-based starting points; we have not generated or auditioned their audio and do not present them as a benchmark.

What Drama 3 does in a short-film workflow

Drama 3 is Fish Audio's text-to-speech model, not a video model. Fish Audio's changelog dates its announcement to October 13, 2026. Its product page currently labels it Preview: takes can vary. Treat the first result as an audition, not the finished soundtrack.

The practical deliverable here is one accepted dialogue scene plus a record of which voice plays each character. It is not a lip-synced film, an isolated track for every actor or a precisely timed performance.

1. Make a cast sheet before adding tags

Use this original scene, The Last Train. A station attendant finds a passenger holding a ticket to a station that closed years ago.

The identity column belongs to the voice choice; the intention column belongs to the performance direction. Do not recast Jun just because his last line needs more urgency.

Start with the unmarked dialogue:

Mara: The last train left ten minutes ago.
Jun: Then why is someone still calling my name?
Mara: Which platform?
Jun: The one you boarded from. Twenty years ago.

Keep the setting and camera instructions in a separate shot plan. Putting “wide shot of an empty station” into spoken text risks making it part of the narration.

2. Direct the turn, then the important words

The Drama 3 guide documents bracketed directions and sounds, plus paired tags that affect selected words. Here is a restrained first pass. Role names are reading labels, not API speaker identifiers.

Mara: [matter-of-fact, finishing a routine shift] The last train left ten minutes ago.
Jun: [quiet, trying not to sound frightened] Then why is someone still calling my name?
Mara: [concern replacing impatience] Which platform?
Jun: [steady, asking her to remember] The one <emphasis>you</emphasis> boarded from. [pause] <whisper>Twenty years ago.</whisper>

Each cue has a job: Mara changes after hearing Jun's question; “you” changes who the accusation points at; the pause delays the reveal. “Scary, cinematic, emotional, dramatic” would give you less specific decisions to evaluate.

For revision, use a three-part note: what to keep → what to change → where. For example: “Keep Jun's restrained tone; make the accusation clearer; change only ‘you’.” This is our editorial method, not a special model command.

Do not begin by stacking crying, shouting and whispering on the same sentence. First check that the actors and words are right. Then change one intention or one phrase and audition again. A bracketed pause is a performance cue, not a promise of an exact number of milliseconds.

3. Generate a scene in the Fish Audio web editor

Use your own Fish Audio account and check the access and usage allowance shown there.

1. Open the official Drama 3 page and follow its Text to Speech entry. Confirm that the selected model is Drama 3.
2. Add two speakers and assign a voice to each. First audition one plain line per role. Choose voices you can distinguish without looking at the speaker label.
3. Put each turn in the corresponding speaker's dialogue entry. The Mara: and Jun: prefixes above are for the reader; do not add those names as spoken dialogue.
4. Apply the directed lines above. If you use the editor's Auto Tag option, review its additions against your cast sheet rather than accepting a different intention for the scene.
5. Generate a take, listen through the reveal and keep notes using the checks below. Save the selected voice identities and exact input alongside the accepted audio.

The workflow above follows Fish Audio's published editor description; we have not validated the signed-in editor for your account. If the model is unavailable there, check Fish Audio's current access information rather than substituting a different model and calling the result Drama 3.

4. Optional: send the same scene through the API

The TTS API reference maps the first ID in reference_id to <|speaker:0|> and the second to <|speaker:1|>. Keep that mapping unchanged across revisions.

Save this as scene.json. Replace the two placeholder IDs with voice IDs available to your Fish Audio account: Mara first, Jun second. These placeholders are not usable voice IDs.

{
  "text": "<|speaker:0|>[matter-of-fact, finishing a routine shift] The last train left ten minutes ago.\n<|speaker:1|>[quiet, trying not to sound frightened] Then why is someone still calling my name?\n<|speaker:0|>[concern replacing impatience] Which platform?\n<|speaker:1|>[steady, asking her to remember] The one <emphasis>you</emphasis> boarded from. [pause] <whisper>Twenty years ago.</whisper>",
  "reference_id": ["REPLACE_WITH_MARA_VOICE_ID", "REPLACE_WITH_JUN_VOICE_ID"],
  "format": "mp3"
}

With your API key already set in the FISH_API_KEY environment variable, run:

: "${FISH_API_KEY:?Set FISH_API_KEY locally before running}"
curl --fail --show-error --silent \
  https://api.fish.audio/v1/tts \
  --header "Authorization: Bearer ${FISH_API_KEY}" \
  --header "Content-Type: application/json" \
  --header "model: drama-3" \
  --data-binary @scene.json \
  --output scene.mp3

This makes a billable request against your Fish Audio allowance, not Pencil0 credits. We have checked the example against the documentation but have not sent it. According to the API reference, a missing or unrecognized model header falls back to s2.1-pro; a successful response alone does not establish that you requested Drama 3. For 401, check authentication; for 402, check the account's balance or allowance before retrying.

5. Listen with a decision sheet

Keep the original script beside the audio. Listen once without reading, then once while following every word. The first pass tests whether the scene makes sense; the second catches omissions that a plausible performance can hide.

These are proposed acceptance checks, not reported results. Name takes clearly, for example last-train-en-take-02.mp3, and record script version, voice IDs, model and what changed. Keep both the audio and the exact request for the selected take; rerunning the same text need not reproduce it exactly.

Pronunciation and language problems

For an English name that keeps being misread, check the English phoneme guide: it uses CMU Arpabet rather than IPA. For Chinese polyphonic characters, use the Chinese guide, which specifies tone-number pinyin, one syllable per phoneme-tag pair. Correct the smallest ambiguous word and audition it inside the full sentence.

Changing the language of a direction is different from forcing the spoken language. Drama 3's language-switching tags are documented as experimental. For a bilingual film, approve each language as its own performance; translating the text does not preserve timing or acting automatically.

6. Turn the accepted dialogue into an editable scene

Do not lock a four-second close-up before knowing how long its line and reaction need. After approving the audio, mark the actual start and end of each line in your editor, then make a shot plan around those measurements.

This is an editorial proposal, not an automatically generated timing sheet. In an external video editor, place the accepted scene audio under the assembled picture. Decide whether it replaces generated dialogue or is only a performance reference; playing two versions of the same spoken line together can create doubled speech. Review lip sync separately. A good voice track does not synchronize visible mouths by itself.

Where Pencil0 fits today

Use the AI short drama workflow to develop a story into episodes, and the AI storyboard workflow to plan the framing and action of each beat. Keep the cast sheet beside the character consistency workflow when reviewing shots.

Pencil0's current character-voice path uses Gemini TTS to create a short voice reference, which supported video requests can reuse. That reference is different from a finished Drama 3 dialogue track, and it does not guarantee identical voices across takes. There is currently no Drama 3 integration in Pencil0, and this guide does not establish a supported path for importing Drama 3 audio. If you want that audio in the final film, assemble it in your external editor. Pencil0 is currently invite-only; the linked product pages explain the available entry path.

Two-voice text-to-speech troubleshooting

Why does the wrong character speak a line?

Check the voice-ID order and the tag at the start of that turn together. If Jun's ID is first in reference_id, <|speaker:0|> now selects Jun, even if your notes call that role Mara. Fix the mapping before changing the acting directions. Keep reading labels such as Mara: out of the spoken text.

Why does a successful API request sound like the wrong model?

Check that you sent the exact HTTP header model: drama-3. The API documentation says an omitted or unrecognized value falls back to s2.1-pro. HTTP success is not proof that the intended model was selected. Then compare the exact submitted text with your saved script; do not judge a missing tag from memory.

Can I request exact pauses or automatic lip sync?

Treat the pause cue as acting direction. Measure the accepted audio in your editor and set shot lengths around the actual line and reaction. This workflow does not provide frame-accurate pauses, separate actor stems or automatic lip sync; review those needs in the edit.

Can I bring this finished dialogue into Pencil0?

This guide uses an external editor for the final dialogue track. Pencil0's short Gemini character-voice reference is a different input, and this article does not establish finished Drama 3 audio import. For the picture side of the same scene, the character-reference preparation guide provides a fixed identity record and a three-shot visual check.

Before you make a longer scene

• Approve this small cast and one complete exchange first. Do not expand a failure into a full episode.
• Use voices you have permission to use. If creating a custom voice, Fish Audio's reference-recording guidance recommends clean, single-speaker material; background music is not useful character identity.
• Keep a dialogue master separate from camera directions, subtitles and the version containing generation tags. Subtitles should contain spoken words, not speaker tokens or performance instructions.
• Keep the accepted take, exact input and cast mapping together, then revisit the official documentation if the Preview model changes.

Sources checked October 15, 2026: release history, Drama 3 product and editor overview, Drama 3 syntax, hosted TTS API, English pronunciation, Chinese pronunciation and voice-reference recording. Scripts, cast sheet, shot plan and listening checklist are original editorial examples, not Fish Audio demonstrations or generated results.

Rechecked October 16, 2026: speaker mapping and the model-header fallback against the official Drama 3 guide and TTS API reference.