Resonance

Unified Speech-to-Text (STT) and Text-to-Speech (TTS) API Server.

Features

• STT: GigaAM-v3 model for Russian speech recognition
• TTS: Russian Silero v5 voices and English Kokoro voices
• i18n: Interface available in English, Russian, Chinese

Demo

Quick Start

cp .env.example .env
just build
just run

Open http://localhost:8000

GPU: set DEVICE=cuda in .env before building.

Models are loaded lazily on first real STT/TTS use. Container startup does not pre-download or pre-load model weights, so the first request to a specific backend may take noticeably longer.

Configuration

See .env.example for available options.

API Endpoints

Endpoint	Method	Description
/api/health	GET	Health check
/api/config	GET	Public configuration including TTS language -> voice catalog
/api/models	GET	List backend/model status plus TTS catalog
/api/jobs	GET	List current session jobs (compact DTO); query limit (default 60), offset; JSON includes has_more, next_offset
/api/jobs/stt	POST	Start STT job, returns job_id
/api/jobs/tts	POST	Start TTS job with text, language, voice_id; returns job_id
/api/jobs/{job_id}	GET	Get job status/result (session-scoped)
/api/jobs/{job_id}/events	GET	Stream job events (SSE, session-scoped)
/api/jobs/{job_id}/cancel	POST	Cancel active job (session-scoped)
/api/stream/download	GET	Download TTS audio

F5 Recovery Model

• Frontend stores only active job IDs in localStorage:
  • resonance_stt_active_job_id
  • resonance_tts_active_job_id
• Drawer jobs list is loaded from GET /api/jobs (paginated with offset / has_more) and contains only jobs for the current browser session.
• Backend assigns resonance_session_id cookie and enforces ownership on status/events/cancel endpoints.
• After page reload, UI restores state via GET /api/jobs/{job_id} and continues progress via /events.
• Job data is in-memory on server (JobRegistry), so after server restart unknown job_id is cleared on client and UI resets to neutral state.