zara-voice: runtime TTS and voice-profile plugin #12

New issue

Closed

opened 2026-08-31 18:43:36 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-31 18:43:36 +00:00

(Migrated from github.com)

Plugin

zara-voice

Goal

Own Zara's normal runtime speech synthesis and voice-profile selection without coupling Zara core to one TTS engine.

Scope

enumerate/install/configure runtime TTS backends;
select voice/profile per persona/session;
synthesize text to audio;
stream/cancel playback;
expose supported language/style/emotion controls when the backend provides them;
normalize sample rate/format metadata;
cache generated audio under bounded policy;
expose health/latency/backend state.

Voice creation/cloning/model training belongs in zara-voice-lab, not this plugin.

Backend model

Define a provider-neutral adapter interface suitable for local and remote TTS engines. Local engines should be first-class and secrets must remain outside repository/config output.

Safety / reliability

bounded text length, output duration and cache size;
cancellable synthesis/playback;
explicit backend timeout/error states;
no silent upload of voice/text to a remote provider;
output metadata must identify the backend/profile used.

Acceptance criteria

provider-neutral TTS adapter API exists
Zara can select profile, synthesize, play and cancel
local and remote backend semantics are represented explicitly
emotion/style controls are capability-gated
bounded cache/output policies are tested
tests use fake backends and require no audio device/network
README documents backend/profile configuration
registry validation, plugin tests and nix flake check pass

## Plugin `zara-voice` ## Goal Own Zara's normal runtime speech synthesis and voice-profile selection without coupling Zara core to one TTS engine. ## Scope - enumerate/install/configure runtime TTS backends; - select voice/profile per persona/session; - synthesize text to audio; - stream/cancel playback; - expose supported language/style/emotion controls when the backend provides them; - normalize sample rate/format metadata; - cache generated audio under bounded policy; - expose health/latency/backend state. Voice creation/cloning/model training belongs in `zara-voice-lab`, not this plugin. ## Backend model Define a provider-neutral adapter interface suitable for local and remote TTS engines. Local engines should be first-class and secrets must remain outside repository/config output. ## Safety / reliability - bounded text length, output duration and cache size; - cancellable synthesis/playback; - explicit backend timeout/error states; - no silent upload of voice/text to a remote provider; - output metadata must identify the backend/profile used. ## Acceptance criteria - [ ] provider-neutral TTS adapter API exists - [ ] Zara can select profile, synthesize, play and cancel - [ ] local and remote backend semantics are represented explicitly - [ ] emotion/style controls are capability-gated - [ ] bounded cache/output policies are tested - [ ] tests use fake backends and require no audio device/network - [ ] README documents backend/profile configuration - [ ] registry validation, plugin tests and `nix flake check` pass