Fish Audio Drama 3: give two game characters distinct voices

Give each part one job

Nora and Dex share a coastal weather station, but they have different roles and voices. The application handles their instructions, recent conversation and room state. Fish Audio generates the spoken performance. The final video combines recorded exchanges with a fresh typed-request check in the Unreal scene.

1. Write a short role for each character: Nora welcomes visitors; Dex maintains the equipment.
2. Keep the current lights and door state alongside the conversation.
3. Allow only specific actions, such as switching lights or opening the observation-room door. Validate those actions in your application.

Get one voice request working

Start with one short line before adding speech recognition or multiple characters. Create an API key in your Fish Audio account and keep it on your server. The official Drama 3 documentation specifies POST /v1/tts with the model header set to drama-3; a missing or misspelled model header can select a different model.

1. Choose a suitable voice and retain its reference ID. Use a voice you have permission to use.
2. Send the text “Welcome to the station—and please leave the red switch alone.” with the selected voice reference.
3. Play the returned audio and check the response before connecting the request to the scene. Never include the API key in client files, footage or a public repository.

Direct the same line three ways

Keep the words constant so you can hear the delivery change. The video compares plain delivery, a nervous/reassuring direction and a whisper on part of the warning.

1. Generate the plain sentence as a baseline.
2. Try “[nervous, trying to sound reassuring] Welcome to the station—and please leave the red switch alone.”
3. For a local delivery change, try “Welcome to the station—and <whisper>please leave the red switch alone.</whisper>” and listen to the full result.

Connect dialogue and the room

Send a typed question to the character logic, generate its reply, then send that reply to Fish Audio for speech. Spoken input adds a separate speech-recognition step. The recorded tests used Fish speech recognition and Codex CLI for replies; the later Unreal check used a new typed request. A replay of a recorded-input test does not establish live microphone or interruption support.

1. Ask Nora what visitors should know about the station. Include her role and the current room state with the request.
2. Route a request about the observation-room door to Dex; validate the requested action before changing the scene.
3. Check that the reply, voice and actual door/light state agree. Measure the generation wait rather than treating edited playback as instantaneous.

Keep the two offers separate

Luke’s campaign link offers 10,000 free web-app credits. The separate Drama 3 launch discount applies to API use. The sponsor describes the API offer as $10 per million characters; the current official pricing page meters UTF-8 input bytes and gives an expiry of 12 November 2026 at 17:00 UTC. Check the live pricing before spending.

1. Use the Fish Audio campaign link below for the web-app offer.
2. Read the current Drama 3 and API pricing documentation before running a larger batch.
3. Start with a single short test and inspect its audio, usage and response time before scaling the workflow.

The takeaway

Prove one voice request first, then connect character context and a few validated actions. Judge the spoken result and the scene state together.