ForgeGuard Kokoro Server

ForgeGuard Kokoro Server

ForgeGuard Kokoro Server is a container-native, OpenAI-compatible text-to-speech service built around the Kokoro-82M model. Point any OpenAI speech client at it, deploy it as a container or Helm release, and run natural-sounding speech synthesis on infrastructure you control.

It is designed for operators: the server accepts connections immediately and warms the model in the background, exposes distinct liveness and readiness contracts, ships model weights baked into the images, and offers optional bearer authentication and built-in HTTPS.

Find your path

Get running

• Quickstart — one container, a health check, and your first synthesized clip.
• Container deployment · Docker Compose · Kubernetes (Helm)
• Hardware profiles — supported GPUs and CPU fallback.

Operate safely

• Health and readiness — the exact probe contract.
• Security hardening — authentication, TLS, container posture.
• Observability · Upgrades
• Responsible use — synthetic-media disclosure guidance.

Integrate

• OpenAI-compatible API — /v1/audio/speech, voices, models.
• Extended API — captions, phonemes, inline controls, operational endpoints.
• Configuration · Environment variables
• Compatibility

Understand the system

• Architecture overview · Request lifecycle
• Voices and languages · Streaming and audio formats

What you get

Capability	What it provides
OpenAI-compatible API	A drop-in /v1/audio/speech endpoint any OpenAI SDK can target.
Designed multi-language voices	English (US/GB), Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, and Mandarin Chinese.
Voice mixing and controls	Weighted voice combinations, per-word timestamps, phoneme in/out, and inline pause/pronunciation tokens.
Streaming output	mp3, wav, opus, flac, and raw pcm.
Orchestrator-friendly startup	Immediate liveness, background warmup, and a strict readiness contract.
Self-hosted control	Baked-in weights, offline by default, optional bearer auth, and built-in HTTPS.

Distribution

ForgeGuard Kokoro Server is distributed as container images and a Helm chart only. There is no supported bare-metal install path. Images are published to the GitHub Container Registry under ghcr.io/forgeguard-ai/. latest tracks the newest stable release; pin a release tag (for example 1.1.0) for persistent deployments.

Roadmap

The following are planned, not current, functionality:

• Real-time voice endpoints — low-latency incremental synthesis over a persistent connection (WebSocket/SSE) with barge-in-friendly cancellation.
• More inference backends — AMD (ROCm) and Intel images as hardware becomes available to develop and validate against.

Lineage and license

This is an original ForgeGuard AI project with documented lineage, derived from remsky/Kokoro-FastAPI. It is licensed under Apache-2.0. See the repository LICENSE and NOTICE for full attribution to Kokoro-82M, the kokoro/misaki libraries, and StyleTTS2.