Neutral voice archetypes + layered delivery

name: Voice profile catalog overview: Bundle multiple Piper speakers, add Novolis.Audio.Voice.Profiles with a small catalog of neutral base-voice archetypes (model + delivery temperament only), and keep ATC/radio/phraseology as composable layers in Voice.Atc. todos:

• id: manifest-models

content: Add lessac-low + kristin-medium to NovolisAudioVoiceModelsManifest; fetch models, LFS, MODEL_CARD; run pipeline generate status: completed

• id: sherpa-multi-zip

content: Generalize pack/extract scripts and SherpaOnnx csproj + targets for all manifest model zips status: completed

• id: voice-profiles-pkg

content: Create Novolis.Audio.Voice.Profiles with neutral archetypes only (4–5 ids); no ATC/radio in package status: completed

• id: atc-delivery-split

content: Refactor AtcVoiceProfile into delivery-only Apply (effects + phraseology); remove model/rate ownership from ATC preset status: completed

• id: tests-docs-release

content: Unit tests, composition examples, docs, verify-nuget-only, version bump and GPR publish status: completed isProject: false

Design principle

Base voice profiles describe who is speaking (TTS model + temperament: rate, optional future pitch/noise knobs). They are domain-neutral — not “ATC tower” or “pilot.”

Delivery layers (applied separately) describe how it sounds on the wire: ICAO phraseology, atc-radio DSP, future radio, intercom, etc. Consumers compose layers explicitly.

flowchart LR
  subgraph base [Voice.Profiles — base only]
    Archetype[excitable_female / procedural_male / ...]
    Model[Piper VoiceModel]
    Rate[SpeakingRate]
    Archetype --> Model
    Archetype --> Rate
  end
  subgraph delivery [Voice.Atc — optional layer]
    Phrase[Phraseology]
    FX[AtcRadioEffects]
  end
  base --> Sherpa[SherpaVoiceSynthesizer]
  Sherpa --> delivery
  delivery --> Play[Playback / WAV]

Current state

• `VoiceProfile`: default, atc — atc conflates role + delivery.
• `AtcVoiceProfile.Apply`: sets model, speaking rate, phraseology, and radio effects in one call.
• Single bundled model: en-us-piper-amy.

1. Bundled Piper speakers (unchanged infrastructure)

Add two models to `NovolisAudioVoiceModelsManifest.cs`:

Models are speakers, not personalities. Personalities live in archetype presets.

Same repo/LFS, fetch script, multi-zip Sherpa packaging as before (sections 1–2 of prior plan): generalize `pack-voice-model-archive.ps1`, `Novolis.Audio.Voice.SherpaOnnx.csproj`, `Novolis.Audio.Voice.SherpaOnnx.targets`.

2. Package: Novolis.Audio.Voice.Profiles

References: Novolis.Audio.Voice + abstractions only — no Novolis.Audio.Voice.Atc, no effects.

Types

/// <summary>Neutral base voice: Piper model + synthesis temperament. No DSP or phraseology.</summary>
public sealed record VoiceArchetype(
    VoiceProfile Profile,
    VoiceModelProfile Model,
    float SpeakingRate,
    string Description);

public static class VoiceArchetypeCatalog
{
    public static VoiceArchetype ExcitableFemale { get; }
    public static VoiceArchetype ProceduralMale { get; }
    // ... small fixed set
    public static IReadOnlyList<VoiceArchetype> All { get; }
    public static bool TryGet(string profileId, out VoiceArchetype archetype);
}

public static class VoiceArchetypeApplicator
{
    /// <summary>Configures synthesis only; leaves effects and text normalization unchanged.</summary>
    public static VoiceServiceBuilder Apply(VoiceServiceBuilder builder, VoiceArchetype archetype);
}

Apply sets only:

• VoiceSynthesisOptions.Profile = archetype id (e.g. excitable_female)
• VoiceSynthesisOptions.ModelProfile
• VoiceSynthesisOptions.SpeakingRate

Does not call UseEffects, NormalizeWith, or touch Voice.Atc.

Initial archetypes (small set — tune rates after listening)

Five archetypes give coverage without role-names. IDs are snake_case and stable for serialization/tests.

Optional DI: AddNovolisVoiceArchetypes() registers catalog only (not IVoiceService — composition stays explicit).

3. Delivery layer: refactor Voice.Atc

Split `AtcVoiceProfile` so ATC is delivery, not base voice:

Remove from ATC apply path:

• VoiceModelCatalog.DefaultProfile
• Default SpeakingRate (1.14) — consumers pick rate via archetype or explicit options

Keep VoiceProfile id atc only if needed for telemetry/tagging when delivery is active; alternatively use a separate VoiceDeliveryTag — prefer leaving VoiceSynthesisOptions.Profile as the archetype id and tagging delivery via options (see below).

Optional: explicit delivery options on builder

Lightweight extension in Novolis.Audio.Voice (or stay in Atc package):

public sealed class VoiceDeliveryOptions
{
    public string EffectChainId { get; init; } = "none"; // atc-radio, none, future radio
    public Func<string, string>? NormalizeText { get; init; }
}

AtcVoiceProfile.ApplyDelivery maps AtcVoiceOptions → effects + phraseology. Sample rate for filters comes from current VoiceSynthesisOptions.ModelProfile via VoiceModelCatalog.TryGet(...).SampleRateHz at apply time (not hard-coded 16 kHz).

4. Consumer composition (documented pattern)

using Novolis.Audio.Voice;
using Novolis.Audio.Voice.Atc;
using Novolis.Audio.Voice.Profiles;

IVoiceService tower = VoiceArchetypeApplicator
    .Apply(new VoiceServiceBuilder(), VoiceArchetypeCatalog.ExcitableFemale)
    .Let(b => AtcVoiceProfile.ApplyDelivery(b, new AtcVoiceOptions { SpeakingRate = 1.14f })) // rate override optional
    .BuildService();

// Dry briefing — base voice only, no radio
IVoiceService briefing = VoiceArchetypeApplicator
    .Apply(new VoiceServiceBuilder(), VoiceArchetypeCatalog.ProceduralMale)
    .BuildService();

Note: if SpeakingRate should live only on archetype, AtcVoiceOptions.SpeakingRate becomes optional override or is removed from ATC options over time (breaking change — defer to follow-up; initially ATC can still accept rate as override for dogfooding).

Dogfooding follow-up: BridgeCommander picks archetype per character, adds ApplyDelivery for comms lines only.

5. Tests, docs, release

Tests

• VoiceArchetypeCatalog lists 5 ids; TryGet round-trip.
• VoiceArchetypeApplicator.Apply sets model + rate; does not register effects (assert identity pipeline or inspect options).
• AtcVoiceProfile.ApplyDelivery adds phraseology + radio without changing ModelProfile when builder pre-configured with archetype.
• Model catalog + conditional Sherpa synthesis per bundled model (unchanged).

Docs

• Profiles README: archetype table + composition with ATC.
• docs/getting-started.md: two-step base + delivery example.
• docs/voice-models.md: speaker table vs archetype table (clarify distinction).

Governance: verify-nuget-only.ps1, GPR publish Voice.SherpaOnnx + Voice.Profiles, version bump.

Package responsibilities (revised)

Risks