Speech services let a personal session send audio for transcription and request spoken output. The current implementation uses Whisper for ordinary transcription and IndexTTS for synthesis. Devices consume the service contract rather than carrying another copy of every server model.
One renderer, a visible queue
Synthesis shares a render lock, bounded queue and model lifecycle. Background preparation yields to foreground requests. Health reports readiness and queue state; it does not claim a listener has accepted the voice or a microphone has captured usable speech.
Ordinary speech passes through a spoken-text filter so markup, code and internal reasoning do not automatically become audio. Private reference-based synthesis has a separate credential and input path. A key for ordinary speech does not grant reference upload.
Keep provenance attached
A machine transcript is an interpretation of captured audio. Language detection, translation and synthesized speech should retain their own origins. The wider cross-language and agent-to-agent delivery design remains separate from the service that can currently render a file.
What remains open
- Endpoint and credential delivery, microphone quality, playback and voice likeness require their own acceptance.
- Cross-language enrollment and capability-controlled agent-to-agent speech remain unfinished.
- Model, service and dependency licensing must be checked independently; this repository has no single declared license.