Repository navigation
Releases
AnyAudio v0.4.2
AnyAudio v0.4.2
这一版修复唱歌环境安装与声库安装报 tar exited with code 1 的问题:子进程输出不再通过管道捕获,避开 DSH 沙箱限制;失败时保留完整错误原因。
中文
修复
- 唱歌安装共用的子进程函数改为把 stdout/stderr 重定向到临时文件,不再用管道捕获。DSH 沙箱会拦截管道捕获子进程输出(EPERM),导致 tar 一启动就被拒、却只报
code 1;改用临时文件后不再触发该限制。 - 子进程失败时,把 tar/pip 写到 stderr 的完整原因拼进错误信息(例如
tar exited with code 1: <真实原因>),不再只有一句退出码掩盖真实原因。 - 改动只涉及唱歌安装,未触碰 ffmpeg、edge TTS、显卡组件等其它子进程调用。
安装
dsh plugin --profile desktop add dsh-any-audio@0.4.2装完重启 DSH。需要 Node.js 22.19+(22.x)或 24+。
English
Fixes
- The shared child-process helper for singing installs now redirects stdout/stderr to temporary files instead of capturing them through pipes. DSH's sandbox blocks pipe-captured child output (EPERM), which made tar fail instantly while only reporting
code 1; temp files avoid that path. - On failure, the full stderr from tar/pip is now included in the error (for example
tar exited with code 1: <real cause>) instead of a bare exit code hiding the real reason. - The change is limited to singing installs; ffmpeg, edge TTS and the GPU component are untouched.
Install
dsh plugin --profile desktop add dsh-any-audio@0.4.2Restart DSH afterwards. Requires Node.js 22.19+ (22.x) or 24+.
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.
AnyAudio v0.4.1
AnyAudio v0.4.1
这一版修复音源下载体验:需要手动下载的特色音源不再弹出必然失败的报错,改为直接打开官方下载;电吉他音源指向完整版 SoundFont 压缩包直链。
中文
修复
- 设置页「推荐音源」里需要手动下载的音源(如 Electric Guitar Clean、Upright Piano KW),按钮从「下载安装」改为「下载」,点击直接打开官方下载页,不再弹出 "has to be downloaded manually" 的报错。
- 展开「详情」时增加一行安装说明:解压后把 .sf2 放到共享音源目录,并写明应保持的文件名。
- Electric Guitar Clean 的下载地址改为完整版 SF2 压缩包(EGuitarFSBS-clean-SF2-20260807.7z)的直链,点击即下载,不再落到要在 8 个文件里自己挑的 Release 页面。
- 补齐唱歌步骤一处缺失的英文翻译。
安装
dsh plugin --profile desktop add dsh-any-audio@0.4.1装完重启 DSH。需要 Node.js 22.19+(22.x)或 24+。
English
Fixes
- For manual-download banks (Electric Guitar Clean, Upright Piano KW), the settings button changes from "Install" to "Download" and opens the official download page instead of throwing a "has to be downloaded manually" error.
- Details now include an install hint: unpack the .sf2 into the shared font folder, keeping the exact file name.
- Electric Guitar Clean now links directly to the full SF2 archive (EGuitarFSBS-clean-SF2-20260807.7z) instead of a release page with eight files to choose from.
- Added a missing English translation in the singing step.
Install
dsh plugin --profile desktop add dsh-any-audio@0.4.1Restart DSH afterwards. Requires Node.js 22.19+ (22.x) or 24+.
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.
AnyAudio v0.4.0
AnyAudio v0.4.0
这一版加入了唱歌:Agent 按音符和歌词在本机合成人声,作为单独一轨进剪辑台。运行环境和声库按需下载,不打包进插件。
中文
唱歌
新增第 6 个工具 anyaudio_sing。音符四字段:[起始秒, 时长, 音高或 null, 歌词]。按休止自动分句、逐句合成;一个字唱多个音时,字写在第一个音符上,后面的音符歌词留空。合成结果落在独立人声轨上,可试听、调音量、逐句重生成。Agent 遇到唱歌需求时自动调用,无需点名。
本机合成用 DiffSinger 和绮萱声库,全程不联网。首次使用分三步:装运行环境(Python 和依赖,约 45 MB,声码器约 50 MB);从 GitHub、百度网盘或夸克网盘下载声库 zip(三处是同一个原包);在设置页填本地 zip 路径安装。显卡加速是可选增强,需额外下载约 216 MB 的组件。
实测:3 分钟的歌,CPU 约 247 秒,显卡约 28.5 秒,快约 8.7 倍。
引导与设置
首装引导从五步变六步,新增「装唱歌引擎」,放在「认识剪辑台」之前。设置页新增「唱歌」标签,可查看安装状态、选声库和推理设备、看下载进度。系统提示词和两个技能文档加入自动触发说明。
已知限制
音质由声库和调教决定,插件只提供安装渠道、接口和使用方式。唱歌声库只保证绮萱 v2.7.0 基本可用;其它声库接口开放,但不保证加载和效果。显卡加速依赖本机显卡与驱动,不兼容时重装唱歌环境即可回到 CPU。
安装
dsh plugin --profile desktop add dsh-any-audio@0.4.0装完重启 DSH。需要 Node.js 22.19+(22.x)或 24+。
English
Singing
A sixth tool, anyaudio_sing, synthesises a real vocal on this machine from notes and lyrics and places it on its own track in the clip bay. A note is [startSec, durSec, midi|null, lyric]. Notes are split at rests and sung in sequence; for a drawn-out syllable, put the word on the first note and leave the rest empty. The result can be auditioned, gain-adjusted and regenerated line by line. The agent calls it automatically on singing requests.
Synthesis uses DiffSinger with the Qixuan voicebank, entirely offline. First use takes three steps: install the runtime (Python and dependencies, about 45 MB, plus a 50 MB vocoder); download the voicebank zip from GitHub, Baidu Drive or Quark Drive (the same archive everywhere); then install it from Settings by pasting the local path. GPU acceleration is optional and adds about 216 MB.
Measured: a three-minute song takes about 247 s on CPU and 28.5 s on GPU — roughly 8.7× faster.
Guide and Settings
The first-run guide grows from five steps to six, adding "install the singing engine" before the clip bay. A new Singing tab shows install status, voicebank and device selection, and download progress. The system prompt and both skills now describe the auto-trigger.
Known limitations
Vocal quality is set by the voicebank and tuning; the plugin only provides the install channel, interface and usage. Only Qixuan v2.7.0 is guaranteed to work; other banks are accepted but not guaranteed. GPU acceleration depends on local hardware and drivers; reinstalling the singing environment restores CPU if needed.
Install
dsh plugin --profile desktop add dsh-any-audio@0.4.0Restart DSH afterwards. Requires Node.js 22.19+ (22.x) or 24+.
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.
AnyAudio v0.3.0
AnyAudio v0.3.0
这一版的音效从「合成」换成了「真实录音」:装一包真实音效,agent 按名字检索、看着波形挑选,再摆到时间线上。配音多了第三个引擎,设置页和首装引导也一并重做。
中文
本版核心:音效从合成变成真实录音
0.2.0 的音效只有两个来源:音源自带的少量采样,以及 agent 找不到时手写的 NumPy 合成——后者听起来明显廉价。
0.3.0 引入本地音效库:
- 装一包真实录音(推荐 BaseSFX,237 条,CC0 许可),覆盖环境、天气、门窗、脚步、碰撞、UI、转场;装一次,所有工程共用。
- agent 按名字检索,可以要一张波形图看着形状挑,而不是只凭文字猜。
- 库里没有时,自动去线上找 CC0 音效(Freesound)。
- 两者都无果,才回落到采样或程序化合成——并且会在回复里说明「这是合成的」。
音效的真实感是这一版最大的变化。同时支持自己制作音效包:把文件夹按 FOL(动效)/ AMB(环境音)/ DSN(设计音效)组织好,不需要写任何清单文件,直接安装即可。
DeepSeek TTS:第三个配音引擎
用本机 DSH 登录的 DeepSeek 账号,不需要额外配置任何 Key。两个音色(stella 女声、tide 男声),不支持调速/音高/音量,按字数计费。
Edge、MiMo、DeepSeek 三个引擎共用一套接口,选哪个由用户在设置里定;失败时报错给 agent,由 agent 决定换哪个,工具不替它猜。
设置页统一与首装引导
- 「音源」和「音效」两页现在是一套布局:推荐在上、已安装在中、安装和目录在下,不再一个是卡片一个是分隔线。
- 「配音」页重做成三行:引擎、音色、当前默认,跟着引擎切换参数区。
- 删掉了「诊断」页,有用的内容收进「高级」。
- 新增五步首装引导:欢迎 → 装音源 → 装音效包 → 选配音引擎 → 完成。每一步都能跳过,装完就知道该做什么。老用户升级后也会看到一次(0.2 没有引导页)。
其它改动
- 删掉
anyaudio_mix工具(六个变五个):混音能力并入anyaudio_project的渲染流程,不再需要单独一步。 - 下载失败的提示会跟随界面语言;连不上 GitHub 时,中文用户看到中文说明。
- 基础音效包新增夸克网盘下载地址(百度网盘、夸克网盘、Google Drive)。
- 修复剪辑台两个真实问题:预览调用了没导入的函数(导致播放头不动、人声听不到、暂停停不下来),以及选中轨道头背景半透明导致片段透出来。
安装
dsh plugin --profile desktop add dsh-any-audio@latest装完重启一次 DSH。需要 Node.js 22.19+(22.x)或 24+。
升级后打开新对话会看到引导;已有工程继续可用(单个工程上限 15 分钟、20 轨)。
English
The headline: sound effects now come from real recordings
In 0.2.0 an effect could only come from the few samples inside a SoundFont bank, or — when the agent found nothing — from hand-written NumPy synthesis, which sounds cheap.
0.3.0 adds a local sound library:
- Install a pack of real recordings (the recommended BaseSFX holds 237 sounds under CC0) covering ambience, weather, doors and windows, footsteps, impacts, UI and transitions. Install once, shared by every project.
- The agent searches by name and can ask for a waveform picture to pick by shape instead of guessing from words.
- When the library has nothing, it reaches for online CC0 effects (Freesound).
- Only when both fail does it fall back to samples or procedural synthesis — and it says in the reply that the effect is synthesised.
How real the effects sound is the biggest change in this release. Making your own pack is supported too: organise a folder into FOL (foley), AMB (ambience) and DSN (designed sound), and install it — no manifest to write.
DeepSeek TTS, a third narration engine
Uses the DeepSeek account DSH is already signed in with, so there is no key to configure. Two voices (stella female, tide male), no rate, pitch or volume control, billed per character.
Edge, MiMo and DeepSeek share one interface. The user picks the default in Settings; a failure is reported to the agent, which decides whether to switch engines — the tool does not guess for it.
One layout for Settings, and a first-run guide
- Sound sources and sound packs now share one layout: recommendations first, installed ones next, install and folder last. No more one page of cards and one page of divider rows.
- The narration page became three rows — engine, voice, current default — with the parameter block changing per engine.
- The diagnostics page is gone; what was useful lives under Advanced.
- A new five-step first-run guide: welcome → sound source → sound pack → narration engine → done. Every step can be skipped. Anyone upgrading from 0.2 sees it once too, since 0.2 shipped none.
Other changes
anyaudio_mixis removed (six tools become five): mixing folded intoanyaudio_project's render, so it is no longer a separate step.- A failed download is reported in the interface language.
- The base pack gained a Quark netdisk link alongside Baidu and Google Drive.
- Two real clip bay bugs fixed: the preview called a function it never imported (the playhead froze, the voice was inaudible, pause did not stop), and the selected track head was semi-transparent so its clips showed through.
Install
dsh plugin --profile desktop add dsh-any-audio@latestRestart DSH once afterwards. Requires Node.js 22.19+ (22.x) or 24+.
After upgrading, open a new conversation to see the guide. Existing projects keep working; a project is limited to 15 minutes and 20 tracks.
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.
AnyAudio v0.2.0
AnyAudio v0.2.0
This release adds the DAW. Every sound AnyAudio makes now lands on one visible timeline, one track per thing, and can be heard, moved, balanced and re-rendered without asking the agent again. It also changes what the tools produce: audio goes into a project, not into a loose file.
What's in this release
- The DAW. A tab next to the conversation shows the project as tracks of clips. Play from any point, drag a clip, trim it, set its gain, pan, fades and effects, mute or solo, undo. Preview and export share one code path, so what you hear is what renders.
- Seven effects on clips, tracks and the bus: EQ, compressor, limiter, reverb, delay, sidechain and pitch, with one implementation used by both the browser preview and the offline renderer.
- Generated audio becomes editable material. Music, narration and effects are written into the project as separate clips, one per take and per stem, so a single line or instrument can be re-heard, moved or re-rendered on its own. Effects are split by category and by overlap instead of being baked into one file.
- Every material is measured. Each render reports per-track RMS, and the measured level is stored next to the file, so balancing never means guessing and never means rendering again.
anyaudio_project, a sixth tool: create, add material, read, update and render a project. Reading returns track ids and clip gains, so the agent can act on what is actually there.- Hand-off in both directions. Selecting a range and choosing "add to conversation" puts a reference in the prompt; equally, you can adjust one clip until it sounds right and ask the agent to apply that to the rest.
- The DAW follows the interface language and colour, survives a tab switch, keeps unsaved work in a draft, and names its own measures (RMS vs true peak, chosen source vs fallback).
Fixes
- A solo or mute pressed while paused is heard on the next play.
- A seek during playback moves the sound, not just the clock.
- The playhead follows the real head width instead of a constant.
- Playing at the end of the project rewinds instead of doing nothing.
- Stems are scratch files, not deliverables; generated audio stages in scratch space instead of
out/. - A clip change finds its own track instead of demanding an id.
- An empty effect chain is no longer written to
project.json. - An unknown instrument now names its nearest neighbours.
- A voice is matched to the engine that owns it.
Install
dsh plugin --profile desktop add dsh-any-audio@latestRestart DSH once afterwards. Requires Node.js 22.19+ (22.x) or 24+.
Existing projects keep working; a project is limited to 15 minutes and 20 tracks.
Links
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.
AnyAudio v0.1.1
AnyAudio v0.1.1
Patch release. Fixes a bug that could make anyaudio_sources return nothing at all, plus smaller correctness fixes and a batch of documentation gaps the agent needed.
Fixes
anyaudio_sourcesno longer loses its whole result. When no default sound source had been saved yet, the tool returneddefaultSource: nullwhile its output schema declares a string. The host rejected that as aToolOutputErrorand discarded the entire payload, so the agent saw only an error instead of the installed banks, their family scores and the search folders — and went looking for the answer in the plugin source. The key is now omitted when unset.- A score can be passed as a
.jsonfile.anyaudio_music(andanyaudio_sfxforitems) accept an object, a JSON text or a path to a file. Relative paths resolve against your project, not the host process's working directory. Long scores no longer have to be inlined into the tool call. - The reported plugin version comes from
package.jsoninstead of a hard-coded string, so the settings page matches the installed release. - Regression test for output schemas. The suite now validates every tool's real return value against its declared output schema, so this class of bug fails in CI instead of in the agent's hands (40 tests, up from 39).
Documentation
- The 16-track limit per score is now stated in the
anyaudio-scoringskill, along with the file-path form and the headless renderer. scripts/anyaudio-cli.mjs(render a score without DSH) is documented in both READMEs.- New hero banner at the top of the README.
Install
dsh plugin --profile desktop add dsh-any-audioRestart DSH once afterwards. Requires Node.js 22.19+ (22.x) or 24+.
Links
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.
AnyAudio v0.1.0
AnyAudio v0.1.0
First stable release. AnyAudio is an audio production toolkit for AI agents running on DeepSeek Harness: the agent describes the sound, the plugin renders real instruments, speech, effects and the final mix.
What's in this release
- Five tools for the agent:
anyaudio_music,anyaudio_tts,anyaudio_sfx,anyaudio_mix,anyaudio_sources, plus the bundledanyaudio-scoringskill that documents the score format and mixing workflow. - Real instrument timbres rendered offline from local SoundFont banks through spessasynth_core. No SoundFont is bundled — install GeneralUser GS, MuseScore General, Upright Piano or Electric Guitar Clean from the settings page (or
npm run font). The library is shared by every project and agent session on the machine. - Sample-accurate arrangement — note-level scores on one timeline shared by music, narration and effects, with per-track gain, pan, reverb, chorus, fade and offset.
- Speech through Edge TTS (no key required) and MiMo-V2.5-TTS (your own Open Platform key), 11 Edge voices and 7 MiMo voices.
- Sound effects — 52 preset effects from the installed banks plus 15 procedural generators (wind, rain, footsteps, thunder and so on).
- Mix and mux — narration ducking, peak normalisation, and video muxing that keeps the picture's own length: a shorter mix is padded with silence, a longer one is cut at the last frame.
- Settings page in DSH for the sound library, TTS defaults and mixing defaults, with download progress that survives closing the panel.
Install
dsh plugin --profile desktop add dsh-any-audioRestart DSH once afterwards, then open the AnyAudio page in DSH Settings to install a sound source.
Requires Node.js 22.19+ (22.x) or 24+. ffmpeg is only needed for speech and for muxing video.
Links
- Source and documentation: https://github.com/cloudoyx/any-audio
- npm: https://www.npmjs.com/package/dsh-any-audio
- MIT licensed.