Journal · Sound

Every voice, every footstep.

A free game has to sound as good as a paid one, or players feel the difference in the first ten seconds. Here is how a small foundation gives its apps a full cast, an original score and a world of sound, with ElevenLabs doing work that would once have needed a studio, a casting director and a composer.

App Studio makes apps and games that are free for everyone: no price, no ads, no in-app purchases. That promise only means something if the apps are good. Sound is where small teams usually fall short, because voice actors, composers and sound designers are expensive, and every change to a script means another session. ElevenLabs changed that arithmetic for us. Four of our apps now carry sound made with it, and each one uses it differently.

A cast of 118, in two languages

Charlie: Save the Puppy is a story-driven comedy game on iPhone, iPad, Mac, Apple TV and Apple Watch, starring a mini Australian Shepherd loose on the Upper East Side. It talks a lot. Every line is voiced, from the narrator’s nature-documentary commentary to the smallest walk-on part.

Speaking roles, each with its own voice settings
118
Distinct ElevenLabs voices across the cast
59
Voiced clips shipped, every wording of every line
29,977
Ambience loops and sound effects recorded
322

We use Eleven Multilingual v2, which our script calls “the most dependable expressive model”. Multilingual matters here for a specific reason: one character, the Maid, speaks Spanish, and with a multilingual model and a native Spanish library voice her Spanish is native too, not English with an accent painted on.

Every voice comes from ElevenLabs’ premade voices or its public voice library. Our rule, written into the project’s README, is to use stock voices only and never clones of real people. A game about real New York neighborhoods could easily drift into imitation; this keeps it on the right side of that line.

Saying names the way people say them

Multilingual v2 doesn’t take phoneme tags, so when it mispronounced a name we didn’t fight the model; we changed what it was given. A small table rewrites only the text sent to the API, while the subtitles on screen stay as written:

// eleven_multilingual_v2 takes no phoneme tags, so we respell what it reads
const SAY_AS = [[/\bNichole\b/gi, "Nicole"], [/\bAunty\b/gi, "Auntie"], [/\bFederico\b/gi, "Federíco"]];

“Federíco” puts the stress on the third syllable, where an Italian speaker puts it, and the model follows the accent mark.

Never paying twice for the same line

A game script changes constantly, and re-recording everything on each change would waste both money and time. Each clip’s file name is a hash of the voice, the model and the exact words spoken. If a line hasn’t changed, its file already exists and nothing is sent. Each clip also gets a fixed seed taken from that hash, so a regenerated line comes back sounding as close to the original as possible.

// scripts/voices.ts (abridged)
const fileHash = (voiceId, text) =>
  sha256(voiceId + "\0" + MODEL + "\0" + text).slice(0, 20);

const body = {
  text: sayAs(line.text),
  model_id: "eleven_multilingual_v2",
  seed: parseInt(hash.slice(0, 8), 16) % 4294967295,
  voice_settings: { stability: voice.stability ?? 0.5, similarity_boost: 0.8,
                    style: voice.style ?? 0.15, use_speaker_boost: true },
};
await fetch(`https://api.elevenlabs.io/v1/text-to-speech/${voice.id}?output_format=mp3_44100_64`, …);

The masters stay on our machines. The game ships compact HE-AAC copies made with Apple’s own afconvert, so we can re-encode at any time without generating again. Before a run spends anything, the script reads the account’s remaining balance from the subscription endpoint.

Speech bubbles stay on screen as subtitles, and voices can be switched off. Sound adds to the game; it is never the only way to follow it.

An original score, and a city that breathes

Charlie’s soundtrack, 10 loops and 4 stingers, was composed with Eleven Music from detailed composition plans. Every prompt carries the same two rules: instrumental only, and never name or imitate an artist.

The streets, apartments and parks have their own ambience from ElevenLabs Sound Effects, written as 79 loop prompts and 243 sound prompts. Background sound has a subtle problem: a crowd murmur or a radio can contain words a player half-hears and then can’t place. So we ran the ambience through Scribe, ElevenLabs’ speech-to-text model, listening for intelligible speech. One radio clip had some and was made again.

Cheers in Armenian and Persian

Bonluma, our match-three puzzle on iPhone, iPad, Mac, Apple TV and Android, was made to be played by everyone, including people who use VoiceOver, a keyboard or a game controller. It works fully offline, with no accounts and no data collection, and it is the app where sound and accessibility meet most closely.

Pieces of music, from Eleven Music composition plans
62
Sound effects and ambience
71
Voice cheers and lines, in three languages
39
Network calls to ElevenLabs from the app
0

Bonluma’s cheers for a big cascade are in English, Armenian and Persian. They use Eleven v3, whose audio tags let us ask for delivery as well as words, with the language set for each one:

# tools/audio/generate.py (abridged)
CHEERS = {
    "en": [...],
    "hy": ["Ապրե՛ս", "Համո՜վ", "Հրաշալի՜", "Սքանչելի՜", "Ֆանտաստիկ է"],
    "fa": ["آفرین!", "شیرین!", "عالیه!", "فوق‌العاده!", "محشره!"],
}
{"text": f"[excited] {word}", "model_id": "eleven_v3", "language_code": lang, ...}

Armenian matters to us: the Foundation was created in memory of Rouben Malian, and supporting the Armenian community is part of its work. A celebratory voice in Eastern Armenian, in a free game, is a small thing that very few commercial games would ever make.

Two accessibility rules shape how the cheers play. Each one appears with the same word on screen, so the moment is never sound-only. And a cheer never plays while VoiceOver or Bonluma’s narration is speaking, because a screen-reader user needs that channel clear. (Bonluma’s optional spoken narration uses Apple’s built-in speech.)

Every file is generated before release and bundled with the app. In the README’s words: “The app never contacts ElevenLabs.” That is what lets Bonluma promise no tracking and no data collection, and play on a plane.

Quiet sounds for a robot dog’s diary

Pawprint: Robot Dog Diary is a companion app for families with a Sony aibo. Its sound is deliberately small: 14 pieces from ElevenLabs Sound Effects, six soft looping beds under its Stage scenes and Story Theatre, and eight one-shots, including the paw chime its notifications make.

A narrator for twelve explainers

Admit Path, the Foundation’s educational guidance platform, has an Apple TV app with twelve narrated explainer videos, each two to three and a half minutes long. Here the ElevenLabs work is about clarity rather than character.

  1. Choosing the voice by measurement. The narrator, the premade voice Bella, was chosen from six candidates by how accurately a transcription model understood each one, and by its pauses and pitch movement.
  2. Timestamps drive everything. The with-timestamps endpoint returns the timing of every character spoken. Those timings place each slide and write the English captions, so captions and pictures land exactly on the words.
  3. Numbers read the way people say them. A per-script map changes only what the narrator reads (for example, a dollar figure), while the captions keep the written form.
  4. Music that ducks under the voice. The Eleven Music bed sits about 20 dB below the narration, with the volume automated on the same speech timings.
  5. Checked before release. Every take is transcribed by a second model and compared with the script before it is used.

What comes next: Voice and Sight, and games for blind players

The work above is about apps. The Foundation also has Voice and Sight, its program to help people facing permanent voice loss or visual impairment get free voice technology. [TBC: whether Voice and Sight is taking requests now] [TBC: how Voice and Sight provides access, and with which provider] It can help them keep speaking in their own voice or hear the written world read aloud. If that is you or someone you love, here is how to ask.

App Studio is also developing games for blind and low-vision players, designed around sound and screen readers from the first sketch. Everything we learned voicing Charlie, from pronunciation tables to caching to keeping speech clear of the screen reader, goes into them. If you are blind or have low vision and want to test early versions, tell us.

Built with

  • ElevenLabs Text to Speech: Multilingual v2 for Charlie’s cast and Admit Path’s narrator, Eleven v3 for Bonluma’s cheers.
  • Eleven Music: Charlie’s soundtrack, Bonluma’s score, Admit Path’s music beds.
  • ElevenLabs Sound Effects: ambience, effects and interface sounds in Charlie, Bonluma and Pawprint.
  • ElevenLabs Scribe: checking that background sound carries no words a player could make out.

We make our apps’ voices, music and sound effects with ElevenLabs. The audio is made before release, and none of our apps calls ElevenLabs while people use it. This story describes our own work; ElevenLabs did not write or approve it.

Updated : the apps’ platforms now match their store listings, and ElevenLabs is described as a tool we use.

Make something with us

Sound is half of what makes a game feel finished.

If your company makes tools for voices, sound, music or accessibility, and you want to see them at work in free apps that anyone can download, we would like to hear from you.