Presets

A preset turns this into a finished pack:

python scripts\wvs.py presets list
python scripts\wvs.py quickstart --preset eeyore

It bundles three things: a voice from a provider's licensed catalogue, a delivery direction describing how to speak, and all 43 Waze prompts rewritten in that character's register.

The rewrite is the point. A generic navigation line read in a different voice is a novelty. A line actually written in character is the thing people share.

The shipped presets

They fall into two groups, and the difference is not cosmetic.

Public domain. These rest on an expired copyright term, use a licensed catalogue voice, and can be published without a rights argument.

Tigger comes from the 1928 book, not the 1926 one. He does not appear in the first book at all, and the two entered the US public domain two years apart. That is exactly the kind of distinction a preset's metadata exists to record.

In copyright. These do not rest on an expired term. Each names a third-party community model on Fish Audio that is presented as a clone of the original performance, uploaded by a member of the public without any permission from the rights holder or the performer.

Pooh and Tigger have community models too, and scripts/build_all.py uses them by default so one key covers every character. --catalogue builds those two from their licensed voice instead:

Those two ids live in build_all.py, not in the preset files, and deliberately. pooh and tigger rest on an expired copyright and their presets should keep saying so; routing them through a clone changes only how the lines are spoken, and the run prints that it did.

They are in the repository because people build these anyway, and a pipeline that quietly omits the rights position is worse than one that records it. What the SDK gives you here is an accurate label, not permission. The 43 lines in each are original writing; the character, the name and the voice are not.

python scripts\wvs.py presets show eeyore --lines

Where the character goes, and where it must not

A pack holds one recording per prompt, and Waze plays that same file every time. The nine greetings are the only prompts Waze rotates, picking one at random per drive. Nothing else can vary, so variety has to come from choosing which prompts carry the character.

One turn is heard as a distance callout, often two, then the maneuver, and Waze replays the maneuver clip at each distance. A catchphrase on any of those lands two or three times per junction, every junction. "In a quarter of a mile, if you please" was fine on paper and unbearable on a real drive.

The middle rows are how a drive stays unpredictable without Waze randomising anything: those prompts fire when the road decides, so a forty-minute drive gets several catchphrases at moments nobody can predict, and never the same one twice in a row at a junction. Within the greetings no more than three of nine share an opening. Bugs Bunny is the deliberate exception: every greeting opens "Eh, what's up, doc?" tests/test_repetition.py enforces the plain rows and the greeting rule.

Two speech rates, where a character needs them

Tigger runs at 1.15x for greetings and the reroute, and drops to 1.0x for anything a driver acts on. That split is critical_provider_options in the preset.

The reason is specific: an excited voice at speed makes ExitLeft and ExitRight the likeliest pair in the whole pack to be misheard, and a missed greeting costs nothing where a missed turn costs a junction. The energy survives the slowdown because it comes from the exclamation in the text and from the delivery direction, not from the rate.

A prompt counts as navigation-critical if it carries a required token (a direction, a distance, an exit number) or is heard often enough to matter. That covers maneuvers, distances, ordinals, alerts, arrival, and AndThen; it leaves the nine drive-start greetings and the reroute chime free to be fast.

Presets that need only one rate, like Eeyore and Pooh, simply omit the field.

Where the character goes, and where it does not

Not evenly across the 43 prompts. That is a deliberate design decision with two reasons behind it.

Budget. Maneuver and distance prompts carry the highest weight in the size allocator because they are heard on every instruction. Long lines there cost the most bytes, and the pack has to fit inside roughly 0.8 MB in total.

Repetition. A joke on TurnLeft.mp3 is funny once and grating by the fourth turn of the drive.

So the frequently-heard prompts stay tight and the delivery direction carries the register, while the greetings, arrival, alerts, and roundabout ordinals carry the writing. Eeyore says "Turn left." and the flatness does the work; he says "take the seventh exit. This roundabout is a lot." where you will hear it once a year.

Validation enforces the length half of this: prompts of weight 2.0 or above are capped at 70 characters, everything else at 160.

The rules, and how they are enforced

Three rules are enforced by the schema and the validator, not by asking nicely.

1. A preset cannot carry reference audio

Public domain attaches to a work. It never attaches to a later performance of that work. A 1926 book being public domain says nothing about any recording of any actor reading it, and an audio likeness of a performer carries their own rights regardless of the text's age.

So a preset has no field for reference audio, and the validator rejects any attempt to route one through provider_options:

$ python scripts/wvs.py presets check my-preset
  [error] my-preset: 1 problem(s)
  provider_options contains voice-cloning key(s): speaker_wav.
  Presets never clone a performance.

Be clear about what this does and does not stop. It stops a preset from carrying a voice sample and cloning it here. It cannot stop a preset naming a provider voice id that is already a clone somebody else made, because a voice id is an opaque string and no validator can hear it. The fish presets above are exactly that case. Rule 2 exists because rule 1 has that hole.

2. The copyright status has to be stated

rights.status is required, and must be public-domain or in-copyright. There is no default, so the flattering answer cannot be the one you get by saying nothing:

$ python scripts/wvs.py presets check my-preset
  [error] my-preset: 1 problem(s)
  rights.status must be one of public-domain, in-copyright; got ''.

An in-copyright preset is a legitimate thing to build, and the validator will pass it. What it will not do is let the question go unanswered. The status shows up in presets list, in preflight, and as a warning at build time, so the last thing you see before spending money is what you are actually building.

3. Every line still has to work as navigation

A driver who has never heard of the character must know exactly what to do. The validator checks that each line still contains what it exists to communicate: "left" in a left turn, "quarter mile" in the quarter-mile callout, "seventh" in the seventh-exit prompt.

Distance callouts are the strictest. Never get cute with a number.

$ python scripts/wvs.py presets check my-preset
  in_quarter_mile: 'In a little while, I should think.' does not contain
  'quarter mile' or 'quarter of a mile'. A driver who has never heard of the
  character still has to know what to do.

Adding a preset

1. Answer the rights questions first

Both of them, in this order. If either answer is no, stop.

• Work level. Is the specific work the character appears in in the public domain? Not the character generally, not the franchise, not an earlier book in the same series. Name the work, the author, the year, and the basis. If you cannot state the basis in one sentence, you do not have one yet.
• Which work. If the character appears across several books, which one are you drawing on, and is that one clear? Tigger and Pooh differ by two years for exactly this reason.
• Performance level. Are you using a catalogue voice with written direction, and cloning nothing? There is no acceptable version of "just a little bit like the film."
• Adaptation level. Is every line drawn from the book character rather than a later adaptation? No catchphrases invented by a studio, no songs, no characters added later, no dialogue from a film.
• Trademark. Are you aware the character's name may be a live trademark even where the text's copyright has expired? Trademark does not expire. This is the weakest part of the whole arrangement and you should know that going in.

• Jurisdiction. Public domain is not global, and you must state which country your basis applies to. For the Milne books specifically:

  Country	Status
  United States	Public domain. 1926 book from 1 Jan 2022; 1928 book from 1 Jan 2024.
  Canada	Public domain since 1 Jan 2007. Milne died 1956 and Canada was life plus 50 then; the 2022 extension to life plus 70 was not retroactive.
  UK and EU	Still in copyright until 1 January 2027 (life plus 70).

  Anywhere else, check before assuming.

2. Write the lines

copy presets\eeyore.json presets\my-preset.json

• All 43 prompts, no exceptions.
• Frequently-heard prompts stay short. Let the delivery direction carry them.
• Distance callouts keep their numbers exactly.
• Read the chained ones aloud: in_half_mile + turn_right, and turn_left + and_then + turn_right. They have to flow.
• Vary the nine StartDrive greetings. They are the cheapest place to be funny.

3. Check it, before spending anything

python scripts\wvs.py preflight --preset my-preset

Pre-flight runs everything that can be checked without an API call: preset validation, the Waze filename mapping, the clarity rules, and an estimated size against the cap. It also prints what it cannot tell you, which is mostly real clip duration and how the lines actually sound.

python scripts\wvs.py presets check my-preset

presets check is the narrower validation-only version, and is what CI runs. It validates the rights block, refuses cloning fields, checks every prompt is present and still unambiguous, and enforces the length limits.

4. Build it and listen

python scripts\wvs.py quickstart --preset my-preset
python scripts\wvs.py qa --route chained_maneuvers

Check the utilisation line. Target is 85% of Waze's cap and the build fails above 92%. The build's number is measured from the encoded files; pre-flight's is an estimate from character counts. If the two differ by more than 10% the build says so, which means the estimate reads wrong for that voice and you should trust the build.

A preset with slow delivery produces longer clips, so confirm your own numbers rather than assuming the shipped ones transfer.

chained_maneuvers is the route to listen to hardest. If anything in the pack sounds spliced, AndThen is where you will hear it.

5. Open the pull request

Include:

• The completed rights checklist above, with your answers, not ticks.
• The utilisation figure from your build.
• A note on which prompts you left plain and why.

A preset that fails presets check will not be merged. A preset that passes but whose rights basis is hand-waved will not be merged either.

Worked example: why Eeyore is written the way it is

The joke is that the directions are perfect and the delivery has given up. That only survives repetition if the maneuvers stay clipped, so:

• turn_left is exactly "Turn left." The register is entirely in the delivery direction: "slowly and flatly with a resigned, gloomy affect... a slight downward inflection ending every phrase."
• arrived gets "You've arrived. Well. That's that." You hear it once.
• start_drive_4 gets "Setting off. Thanks for noticing me." That phrase is from the 1926 book, which is the point: it is drawn from the work in the public domain, not from a later adaptation.
• Nothing in any of the three public-domain presets uses a studio-invented catchphrase. Tigger does not say the thing from the song. That song is from 1968 and is not Milne.

Estimated utilisation sits at 82% for every shipped preset, across 75 to 109 seconds of audio depending on how talkative the character is. That spread landing on the same utilisation is the allocator doing its job, and the case the 85% target exists to absorb.

What was verified, and how

The distance filenames are not guesswork. Eleven real packs were downloaded from the community archive and transcribed offline; 200.mp3 is confirmed as the 0.1 mile callout, 400.mp3 a quarter mile, 800.mp3 half a mile, 1500.mp3 one mile. The full method and results are in waze-import-spike.md, including two things it confirmed by accident: the 43-filename list is exactly right, and real packs sit between 53% and 94% of the size cap.