The Voice Bench

Editorial preview: vendor recordings and hands-on verdicts are pending. What counts as evidence.

AI game character voices: a repeatable audition workflow

The difficult part is keeping the same character through different emotions, new dialogue and late changes. Test direction and repeatability together.

Updated 2026-09-15

The listening bench

Vendor audio is not available in this edition. The labelled demo lets you try the controls; it cannot tell you which product sounds better.

Character dialogue
Same words. Every tool.

You came back. I thought the bridge was gone. Keep your voice down, take the lantern, and follow me. Whatever you hear behind that door, do not answer.

Fish AudioRecording pending
—
ElevenLabsRecording pending
—
0:00 / 0:00

Choose “Try the player demo”, then play a labelled demonstration voice.

What this work demands

Game dialogue is often revised in small batches. A voice needs to survive a changed line, a shout, a whisper and a different emotional state without becoming a different character. Keep stable file identifiers so revisions do not silently overwrite the wrong cue.

The shortlist, before a verdict

Fish Audio and ElevenLabs will be evaluated using the same character passage. Recommendation, runner-up and budget winner remain unassigned until the actual recordings exist. Neither a vendor’s voice library nor our player demonstration is evidence of production consistency.

A practical audition

  1. Write a short scene with calm, urgency and relief. Keep every line identical between vendors.
  2. Use an authorised stock voice or a voice with documented permission. Record restrictions with the character sheet.
  3. Save the model and direction settings alongside every export.
  4. Generate five takes of the hardest line. Count usable takes before selecting a favourite.
  5. Recreate one line in a new session and compare it with the original scene.
  6. Export dry speech when possible. Apply game-specific processing in your audio pipeline so it remains consistent across voices.

The failure to catch early

A dramatic one-off performance can be difficult to recreate. Compare new lines with the character’s existing dialogue before extending the project. Include retakes and manual cleanup in the production budget.

See Fish Audio versus ElevenLabs for the comparison and how we test emotional range for the scoring definition.