Best character voice tools for Codex users

One line, two models

The line below is the kind of thing a game needs hundreds of: a state change said out loud so the player hears it without reading. We sent it to two of the three speech models with no voice chosen and no settings tuned, so what you hear is each model's default.

Both are usable and they are not the same read. MiniMax lands a little brisker and flatter, which suits a status callout; VoxCPM is softer and slightly slower, which suits a companion character. Neither needed a voice picked from a library, and hearing the difference costs almost nothing: each run was $0.006 for 83 characters at $0.002 per 40 characters (text).

The economics of a line list

The point of that table is not that voice is cheap. It is that re-recording is cheap. A line list is never right the first time: the level changes, a mechanic gets renamed, a joke stops being funny, and with human recording each of those means scheduling a pickup session. Here it means running the list again for the price of a sandwich, which is the difference between dialogue you fix and dialogue you live with.

Two practical consequences. Keep the line list as a file in the project rather than as strings scattered through the code, so regenerating the changed lines is a loop rather than an archaeology exercise. And name the output files after the line ID, not after the text, so a wording change does not orphan the audio the engine is looking for.

The third model bills differently

MiniMax TTS and VoxCPM both bill $0.002 per 40 characters (text) of script, so you can price a line list before you run it by counting characters. Seed Audio is $0.01 per second and settles on the seconds of audio it produces, so the amount is known after the run rather than before. For a long list where you want a predictable total, the script-priced models are the easier choice; for a handful of hero lines where the read matters more than the arithmetic, it is worth hearing all three.

Where the voice goes next

1. Straight into the engine

   The mp3 lands in your project. For barks that is the whole pipeline — no editing, no mastering, just the file and the line ID.

2. Into a lip sync run

   The lip sync services render against the audio you supply, so generate and approve the line first. A script change after syncing means paying for the video again, not just the voice.

3. Under a trailer

   A narrator line plus a music bed is two runs and a local mix. The agent can lay them together with ffmpeg for nothing.

4. As a subtitle source

   Run the finished audio back through transcription and you get timed text for accessibility settings without typing timings by hand.

What generated voice is still bad at

Short, functional lines are where this works. Long emotional performances are where it does not: a model reads a sentence, it does not act a scene, and across a hundred lines the sameness of the delivery becomes the thing players notice. Names and invented words are the other weak spot — a made-up place or an unusual character name will be pronounced the way the spelling suggests, which is often not the way you say it in the studio.

Both have cheap workarounds. Break long speeches into shorter lines and let the game's pacing do the acting; the shorter lines also cost less and regenerate independently. And spell unusual names phonetically in the script that goes to the model while keeping the real spelling in the subtitle text — two fields in your line list rather than one, and the mismatch never reaches the player.

Running it from Codex

Codex can hold the line list, name the files, and wire them into the build, but it cannot speak; the Skill installed with the CLI adds the voice generation capability to the same session. Every run is priced before it starts, each one is itemised on the wallet statement, and an empty wallet stops the batch before anything is charged.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.codex/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.codex/skills/actionway"

You need Node.js 20 or newer and an Actionway account. Claude Code users copy the same folder into ~/.claude/skills/actionway instead, and both routes reach the same catalog and the same wallet — the get started page lists the steps per client side by side.