Mynah — 自己服务器上的实时数字人系统。一张显卡、完整控制台、Apache-2.0
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Host-page abilities: let the embedding page lend the model tools over the DataChannel The browser SDK (web/sdk) has advertised ability()/<mynah-ability> since the initial release, but cored never handled the "abilities" message, so a dashboard that declared abilities got no tool calls back. This is the server half of that contract. - core: ToolSpec/ToolCall types and the optional ToolChatProvider widening of ChatProvider (same pattern as ChatStreamDebugger); dify/coze never implement it and keep their exact code path. - internal/llm: StreamChatTools for the openai provider — one round with tools advertised, streamed tool-call fragments reassembled in model order. - internal/session: per-session ability registry; a turn runs up to 3 model rounds. Action abilities answer the model immediately so the page change and the sentence land together; blocking abilities speak a filler, wait up to 6 s, then let the model answer with the result. - internal/signaling: CORS on the visitor endpoints + OPTIONS preflight so an SDK page served from another origin can reach /offer, /human, /channel/*. Origin is reflected, never with credentials (auth is a body token). - tools/probes/toolprobe: headless pion client that checks the two ordering properties unit tests can't see. No abilities declared (every visitor page today) = unchanged path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
wav2lipLS: publish the 256 weights, say where the 384 ones come from; add a contact address Issue #1 pointed out that the "lightest" profile advertised 6 GB cards but WAV2LIPLS_CKPT_URL had no value anyone could fill in. The 256 checkpoint is now a v0.1.0 Release asset and wav2lipls.sh downloads it when FACE_SIZE=256 and no URL is set (256 is now the default size). The 384 checkpoint is distributed through the community group or on request; README, hardware pages, compose notes and LICENSING say so instead of implying it is bundled. Commercial support sections get a real address (mynah@au56.com) in place of the TODO; a commercial issue label now exists to match. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
wav2lipLS: publish the 256 weights, say where the 384 ones come from; add a contact address Issue #1 pointed out that the "lightest" profile advertised 6 GB cards but WAV2LIPLS_CKPT_URL had no value anyone could fill in. The 256 checkpoint is now a v0.1.0 Release asset and wav2lipls.sh downloads it when FACE_SIZE=256 and no URL is set (256 is now the default size). The 384 checkpoint is distributed through the community group or on request; README, hardware pages, compose notes and LICENSING say so instead of implying it is bundled. Commercial support sections get a real address (mynah@au56.com) in place of the TODO; a commercial issue label now exists to match. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Host-page abilities: let the embedding page lend the model tools over the DataChannel The browser SDK (web/sdk) has advertised ability()/<mynah-ability> since the initial release, but cored never handled the "abilities" message, so a dashboard that declared abilities got no tool calls back. This is the server half of that contract. - core: ToolSpec/ToolCall types and the optional ToolChatProvider widening of ChatProvider (same pattern as ChatStreamDebugger); dify/coze never implement it and keep their exact code path. - internal/llm: StreamChatTools for the openai provider — one round with tools advertised, streamed tool-call fragments reassembled in model order. - internal/session: per-session ability registry; a turn runs up to 3 model rounds. Action abilities answer the model immediately so the page change and the sentence land together; blocking abilities speak a filler, wait up to 6 s, then let the model answer with the result. - internal/signaling: CORS on the visitor endpoints + OPTIONS preflight so an SDK page served from another origin can reach /offer, /human, /channel/*. Origin is reflected, never with credentials (auth is a body token). - tools/probes/toolprobe: headless pion client that checks the two ordering properties unit tests can't see. No abilities declared (every visitor page today) = unchanged path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Host-page abilities: let the embedding page lend the model tools over the DataChannel The browser SDK (web/sdk) has advertised ability()/<mynah-ability> since the initial release, but cored never handled the "abilities" message, so a dashboard that declared abilities got no tool calls back. This is the server half of that contract. - core: ToolSpec/ToolCall types and the optional ToolChatProvider widening of ChatProvider (same pattern as ChatStreamDebugger); dify/coze never implement it and keep their exact code path. - internal/llm: StreamChatTools for the openai provider — one round with tools advertised, streamed tool-call fragments reassembled in model order. - internal/session: per-session ability registry; a turn runs up to 3 model rounds. Action abilities answer the model immediately so the page change and the sentence land together; blocking abilities speak a filler, wait up to 6 s, then let the model answer with the result. - internal/signaling: CORS on the visitor endpoints + OPTIONS preflight so an SDK page served from another origin can reach /offer, /human, /channel/*. Origin is reflected, never with credentials (auth is a body token). - tools/probes/toolprobe: headless pion client that checks the two ordering properties unit tests can't see. No abilities declared (every visitor page today) = unchanged path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Host-page abilities: let the embedding page lend the model tools over the DataChannel The browser SDK (web/sdk) has advertised ability()/<mynah-ability> since the initial release, but cored never handled the "abilities" message, so a dashboard that declared abilities got no tool calls back. This is the server half of that contract. - core: ToolSpec/ToolCall types and the optional ToolChatProvider widening of ChatProvider (same pattern as ChatStreamDebugger); dify/coze never implement it and keep their exact code path. - internal/llm: StreamChatTools for the openai provider — one round with tools advertised, streamed tool-call fragments reassembled in model order. - internal/session: per-session ability registry; a turn runs up to 3 model rounds. Action abilities answer the model immediately so the page change and the sentence land together; blocking abilities speak a filler, wait up to 6 s, then let the model answer with the result. - internal/signaling: CORS on the visitor endpoints + OPTIONS preflight so an SDK page served from another origin can reach /offer, /human, /channel/*. Origin is reflected, never with credentials (auth is a body token). - tools/probes/toolprobe: headless pion client that checks the two ordering properties unit tests can't see. No abilities declared (every visitor page today) = unchanged path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
wav2lipLS: publish the 256 weights, say where the 384 ones come from; add a contact address Issue #1 pointed out that the "lightest" profile advertised 6 GB cards but WAV2LIPLS_CKPT_URL had no value anyone could fill in. The 256 checkpoint is now a v0.1.0 Release asset and wav2lipls.sh downloads it when FACE_SIZE=256 and no URL is set (256 is now the default size). The 384 checkpoint is distributed through the community group or on request; README, hardware pages, compose notes and LICENSING say so instead of implying it is bundled. Commercial support sections get a real address (mynah@au56.com) in place of the TODO; a commercial issue label now exists to match. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
wav2lipLS: publish the 256 weights, say where the 384 ones come from; add a contact address Issue #1 pointed out that the "lightest" profile advertised 6 GB cards but WAV2LIPLS_CKPT_URL had no value anyone could fill in. The 256 checkpoint is now a v0.1.0 Release asset and wav2lipls.sh downloads it when FACE_SIZE=256 and no URL is set (256 is now the default size). The 384 checkpoint is distributed through the community group or on request; README, hardware pages, compose notes and LICENSING say so instead of implying it is bundled. Commercial support sections get a real address (mynah@au56.com) in place of the TODO; a commercial issue label now exists to match. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
wav2lipLS: publish the 256 weights, say where the 384 ones come from; add a contact address Issue #1 pointed out that the "lightest" profile advertised 6 GB cards but WAV2LIPLS_CKPT_URL had no value anyone could fill in. The 256 checkpoint is now a v0.1.0 Release asset and wav2lipls.sh downloads it when FACE_SIZE=256 and no URL is set (256 is now the default size). The 384 checkpoint is distributed through the community group or on request; README, hardware pages, compose notes and LICENSING say so instead of implying it is bundled. Commercial support sections get a real address (mynah@au56.com) in place of the TODO; a commercial issue label now exists to match. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 | |
Host-page abilities: let the embedding page lend the model tools over the DataChannel The browser SDK (web/sdk) has advertised ability()/<mynah-ability> since the initial release, but cored never handled the "abilities" message, so a dashboard that declared abilities got no tool calls back. This is the server half of that contract. - core: ToolSpec/ToolCall types and the optional ToolChatProvider widening of ChatProvider (same pattern as ChatStreamDebugger); dify/coze never implement it and keep their exact code path. - internal/llm: StreamChatTools for the openai provider — one round with tools advertised, streamed tool-call fragments reassembled in model order. - internal/session: per-session ability registry; a turn runs up to 3 model rounds. Action abilities answer the model immediately so the page change and the sentence land together; blocking abilities speak a filler, wait up to 6 s, then let the model answer with the result. - internal/signaling: CORS on the visitor endpoints + OPTIONS preflight so an SDK page served from another origin can reach /offer, /human, /channel/*. Origin is reflected, never with credentials (auth is a body token). - tools/probes/toolprobe: headless pion client that checks the two ordering properties unit tests can't see. No abilities declared (every visitor page today) = unchanged path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 13 天前 | |
Mynah: initial open-source release Open-source digital humans, ready to deploy. Self-hosted realtime digital human: Go realtime core (cored) streaming a lip-synced avatar over WebRTC, pluggable Python engines (MuseTalk 1.5 / FlashHead / wav2lipLS, SenseVoice ASR, Smart Turn end-of-turn), EdgeTTS zero-cost default with Qwen3-TTS and cloud tiers, optional qwen realtime cloud brain, and an admin console for avatars, voices, knowledge bases, channel publishing and engine-pool management. Apache-2.0. Runs on a 6 GB GPU (lightest tier) up to a pooled 24 GB card. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> | 22 天前 |
开源数字人，开箱即用：从引擎到交付，全部开源。
公开演示频道实录：开口提问、唇形同步回答，随时可以插话打断。
FlashHead 引擎 + 内置半身预设 Luna：一张肖像，引擎渲染整头贴回头到腰画布；待机呼吸和害羞、笑、摇头、眨眼等动作在同一画布上烘焙，本能层在大脑组织语言时就选好动作。内置预设，控制台里直接发布成频道。
一张肖像、一条命令，得到一个能实时对话、会做表情、能发布成链接的数字人。 引擎、控制台、交付链路全部开源。文档站：https://honwee.github.io/mynah/zh/
开源、可嵌入的实时数字人引擎 — 全功能开放，无企业版墙。
WebRTC 流式对话数字人：语音进、唇形同步视频出;形象/TTS/ASR/LLM 引擎可插拔, 支持视频训练新形象、声音克隆,一条命令的开箱引导。
硬件门槛先说清楚:默认 FlashHead 形象引擎需要 NVIDIA GPU(4090 级); 轻量 wav2lipLS 路径可跑在更小的卡上。一个形象 worker 同时服务一路会话, 部署规划请先读英文 README 的 Capacity & scaling 章节。完全没有 GPU? 每个组件都可以指向远程 API。
亮点
- 真实时，不是生成视频：云端大脑在你说完 0.2 秒内接话，唇形视频 1 秒内开口，插话立刻停。内置语义轮次模型（Smart Turn v3.2）判断你是否真的说完。数字见英文 README 的 Performance。
- 一张肖像，一个半身形象：FlashHead 渲染整头贴回 720×1280 头到腰画布，待机呼吸与动作片段同画布烘焙，内置 Luna、Nova 预设；自己的肖像一步 prepare。
- 有本能，不只会答：本能层在大脑还在组织语言时选好表情和动作，说完即做。接 Jev 认知模型，断网自动降级到零依赖的关键词规则。见
docs/zh/guide/instinct.md。 - 是控制台，不是脚本：形象、声音、知识库（pgvector RAG + 精排）、频道发布、访问控制、并发配额、引擎池显存账、一键给频道换形象、会话监控。
- 传一段视频做形象：自动选最佳说话基底帧（MediaPipe，不用 InsightFace），烘出活的待机，列表里直接可选。带授权确认。
- 接进你的技术栈：大脑可选本地 LLM、Qwen Realtime 语音到语音，或任意 Pipecat 流水线；形象/TTS/ASR 走 gRPC 与 OpenAI 兼容契约；自带 Python SDK、
pipecat-ai-mynah和 DeepSeek Harness 插件。 - 许可干净：FlashHead（Apache-2.0）、MuseTalk 1.5（MIT）、自训 wav2lipLS 轻量档。见 LICENSING.md。
是什么
cored(Go 实时核心)通过 WebRTC 向浏览器推送音画同步的数字人流:接收
offer → 与 Python 形象引擎建立 gRPC 会话 → 喂入 TTS 音频(文本直驱或 LLM
对话)→ 按 25fps 帧栅格回推 VP8/H.264+Opus。接上 ASR worker 和语义轮次
检测(判断用户"是否真说完了")即是完整的语音对语音闭环。--db 启用管理
控制台:配置热生效、知识库 RAG、可发布的访客频道。
全功能开源:形象训练(传视频→提取最佳说话基底→可选用)、待机烘焙、 声音克隆、动作骨架库、RAG 精排都在这一个 Apache-2.0 仓库里。
快速开始
需要一张 NVIDIA 显卡跑形象引擎（wav2lipLS 6GB，256 版权重公开自动下载、384 版在社区群发放；MuseTalk/FlashHead 12GB，两路同时在线 24GB）、一个 LLM 端点（本地 ollama/vLLM 或任意 OpenAI 兼容 API）和一个 TTS。栈默认用 EdgeTTS 起步，让第一次对话不等 TTS 模型下载；控制台里可热切到 Qwen3-TTS（音色克隆）或云端音色。本机没有显卡时，形象 worker 跑在 GPU 机器上、cored 跑在任何地方即可。
云镜像一键体验（单机体验版）：在 AutoDL 用 Mynah 社区镜像 开一台 24GB 显卡实例，执行 ~/mynah/mynahctl start，再开 SSH 隧道即可，详见云镜像。
最快路径是开箱引导(自动检测/一键安装每个组件,国内网络友好——HF 镜像 + ModelScope):
python3 deploy/setup/helper.py # → http://127.0.0.1:9500
或手动 Docker Compose(见 deploy/compose/README.md):
cd deploy/compose
cp .env.example .env && $EDITOR .env
bash setup-flashhead.sh
docker compose up -d --build
合规提醒
形象训练/声音克隆接口强制 consent 授权字段。克隆真人形象/声音属于 "深度合成",在中国大陆等司法辖区受专门监管——部署前请读 docs/compliance.md。
商业支持
引擎永久免费(Apache-2.0)。需要生产落地帮助的团队,我们提供付费服务:
部署与调优、定制形象/音色、集成开发、托管运行。邮件 mynah@au56.com，
或在 issue 区加 commercial 标签。
更多
架构、API、引擎契约、能力清单见英文 README.md; 安全部署见 SECURITY.md;贡献指南见 CONTRIBUTING.md。
致谢
感谢 LINUX DO 社区佬友们的反馈与支持。