Text to Speech for Unreal Engine: Voice Workflow Fish Audio

Text to speech for Unreal Engine is mostly an asset-pipeline decision. Authored dialogue, combat barks and cutscenes should be generated ahead of time, reviewed and imported as controlled sound assets. Runtime generation is useful for dynamic text, but it adds network latency, moderation, privacy and service-availability problems that a shipped game must handle.

Pre-generated or runtime voice?

Most shipped games should pre-generate every line that has a known script. Runtime generation is the exception, not the default.

How do you prepare dialogue for Unreal?

Write dialogue in a table keyed by character, line ID, locale and trigger. Generate each line as a separate file so a revision does not force the whole scene to be rebuilt. Normalize loudness before importing, and preserve a dry voice that can sit under music and effects.

For a cast, give each character a distinct rhythm rather than simply changing pitch. A fast, clipped support character and a slower narrator are easier to distinguish in combat than two voices with the same cadence.

How do you import generated audio?

1. Download the approved WAV or MP3 output.
2. Normalize and trim silence outside Unreal.
3. Import the file into the project’s content folder using a stable naming convention.
4. Create a sound cue or MetaSound wrapper if the project needs volume, attenuation or randomization.
5. Trigger the asset from Blueprint or the dialogue system.
6. Test the line against compressed game audio and subtitles.

Keep the source file outside the packaged build. Unreal may convert or compress imported assets, and a future re-export needs the original.

How would runtime generation work?

A runtime path needs a trusted backend between Unreal and the speech provider. The game sends text and voice parameters to your service; the service applies moderation, rate limits and authentication, then returns streamed or completed audio. Never ship a private provider key inside a game client.

Runtime generation is appropriate when player-authored text or live content cannot be pre-rendered. Add a timeout, a fallback line and a way to stop audio when the player skips the conversation. Cache repeated phrases so the same greeting is not generated on every encounter.

For speech-layer architecture, read the real-time text to speech guide and voice agent guide. The API overview explains access and concurrency.

How do you localize a game voice cast?

Keep the script in a structured table and generate each locale from the same line IDs. Use a native reviewer for names, cultural references and pronunciation. A single voice can read multiple languages, but that does not mean every delivery or translation will feel natural in every market.

Pre-generated audio is easier to approve than runtime output because localization, legal and audio teams can listen before the build is submitted.

What does a game voice pass cost?

One credit covers about 1,000 characters. Character count, not playtime, determines the initial generation cost.

Alternate takes and pronunciation fixes add to the total. A $29 Creator pack carries 4,000 credits and commercial rights, which is enough for a substantial first game build. The Unity guide covers the same asset-first philosophy in another engine.

What should you test before shipping?

• Every line ID has an asset in every supported locale.
• Dialogue does not overlap unintentionally.
• Subtitles and audio agree after edits.
• Imported volume is consistent with music and effects.
• Runtime requests fail gracefully and never expose a private key.
• A cloned speaker has a signed consent record on file.

For broader casting and batch production, see AI voice for gaming. For character selection, see the AI voice generator guide.

How do you version a game voice set?

Version dialogue by line ID, script revision, voice selection and audio file. Store those values outside the packaged build so a localization pass can replace one file without guessing which take shipped. Keep the original generated file when Unreal imports a compressed asset.

When a scene changes, compare the new line against the surrounding dialogue rather than auditioning it alone. Pace, volume and emotional continuity are properties of the scene. A line can sound good in isolation and break the conversation around it.

For runtime generation, log the text and settings but minimize personally identifiable audio. Ship a known-good fallback line and document what the game does when the service is slow, unavailable or blocked by the player’s network.

Have an accessibility reviewer play the scene with subtitles on and game audio low. They should still be able to follow the turn, identify the speaker and understand any timed objective. If the voice carries information that subtitles omit, the subtitle track is incomplete regardless of how clear the generated audio sounds.

Frequently asked questions

Can I use Fish Audio with Unreal Engine?

Yes. The simplest path is to generate dialogue as audio files and import them as sound assets in Unreal Engine. A runtime integration needs a separate service layer that calls the API and feeds audio into the game.

Should game dialogue be pre-generated or generated at runtime?

Pre-generate authored dialogue, barks and cutscenes. Use runtime generation only when the text is dynamic and the game can tolerate network latency, moderation and service availability constraints.

How do I import generated audio into Unreal Engine?

Download the final file, normalize it, import it as a sound asset, and trigger it through a sound cue or Blueprint. Keep the original script and voice notes with the asset so changes remain reproducible.

How much does a game voice set cost?

One credit covers about 1,000 characters. A 10,000-character dialogue pass costs about 10 credits before alternate takes, so character voice work usually fits in the $29 Creator pack.

Can I ship a cloned voice in a commercial game?

Only with documented voice consent and the appropriate commercial usage rights. The Creator pack and above cover commercial output; consent for the cloned speaker is a separate requirement that must be recorded before production.