Kepos Speech

@lamplitisles/kepos-speech is a DeepSeek Harness Web plugin for Chinese
speech synthesis and short-audio recognition. It adds one optional audio-only
[[tts:text]]...[[/tts:text]] block to a finalized assistant reply, prepares
the resulting MP3, and replaces the block with the browser's native audio
player. The Host service also exposes bounded, synchronous Qwen ASR to trusted
in-process callers.

Locally transcribed user turns use a 🎙️ prefix and can end in a recognized
expression label such as [neutral]. The Speech system prompt defines both as
voice metadata: the marker identifies the source, and the label is an
audio-level ASR classification rather than message content or a claim about the
speaker's inner state.

Source: https://github.com/LamplitIsles/kepos-speech.

Install for DSH

dsh plugin --profile <profile> add @lamplitisles/kepos-speech

The package is pinned to the DSH 0.1.2-rc.1 contract family and contains the
Host entry, browser loader, and cordis.patch.yml bundle patch. The packed
smoke test resolves the official @deepseek-ai/dsh@0.1.2-rc.1 CLI through
npm exec in test-owned temporary state, so it never selects an ambient
dsh. Set DSH_CLI only when intentionally supplying an executable that
reports exactly 0.1.2-rc.1.

Settings and providers

Open the native Kepos Speech Plugin Settings card from a local loopback
DSH Web session and choose Alibaba or ByteDance for tagged TTS output. The
editable Voice IDs default to Maia and
zh_female_sajiaoxuemei_uranus_bigtts; provider-supported IDs up to 128
characters are accepted. The DashScope key is shared by Alibaba TTS and the
fixed Qwen ASR path. The Volcengine key is used only for ByteDance TTS. Both
credential fields are write-only and are stored by DSH as
KEPOS_SPEECH_DASHSCOPE_API_KEY and KEPOS_SPEECH_VOLCENGINE_API_KEY.

The provider selector controls TTS output only. Qwen ASR is the sole
recognition provider and accepts one non-empty supported audio attachment up to
the documented 10 MB encoded bound. It returns complete text and optional
audio-level language and speech-expression annotations; it does not persist
audio or transcript content.

Optional Host service

When mounted, the plugin publishes the optional Cordis service
ctx.get("keposSpeech"). A Host plugin can consume the exported
KeposSpeechService contract:

const speech = ctx.get("keposSpeech");
if (speech) {
  const audio = await speech.synthesize({ sessionId, text }, signal);
  // audio.mediaType === "audio/mpeg"; audio.data is bounded MP3 bytes

  const transcript = await speech.transcribe({
    sessionId,
    mediaType: "audio/ogg",
    data: attachmentBytes,
    language: "zh"
  }, signal);
  // transcript.text plus optional language/expression annotations
}

The service validates the live session and shares the workspace cache with
browser TTS. It is optional, is removed with the plugin lifecycle, and is not a
public transcription or synthesis route. Browser synthesis uses the
authenticated /kepos-speech/synthesize RPC and same-origin
/kepos-speech/audio/... artifacts.

Audio cache

Each prepared passage is keyed by normalized text, provider profile, and cache
format, then written atomically as
.dsh/kepos-speech/audio/<sha256>.mp3 below the active session workspace. A
refresh or remount resolves the session again and reuses the bounded artifact;
there is no browsing, migration, eviction, or cache-management UI.

Maintainer release setup

The tag-only workflow in .github/workflows/release.yml verifies the exact
package that it publishes. Before the first automated release, a maintainer
must bootstrap a distinct prerelease version, then prepare the first stable
version:

1. Create or confirm the @lamplitisles npm scope and manually publish the
   initial @lamplitisles/kepos-speech@0.1.0-beta.0 package so the package
   identity exists. Set package.json to 0.1.0-beta.0, run the bootstrap checks
   below from a maintainer workstation, and publish with local npm
   authentication:

   bun install --frozen-lockfile
   bun run typecheck
   bun run test
   bun run build
   GITHUB_REF_NAME=v0.1.0-beta.0 bun run release:check
   npm publish --access public --tag beta

   This prerelease is deliberately distinct from the later stable 0.1.0;
   do not manually publish 0.1.0.

2. Configure npm Trusted Publishing for
   @lamplitisles/kepos-speech, repository
   LamplitIsles/kepos-speech, workflow
   .github/workflows/release.yml, and the GitHub npm environment.
3. Create the protected GitHub npm environment with the required approval
   policy.

For the first stable trusted publish, push the intended release commit and
create the tag with the supported og operation (og tag --help describes
this as “Create and push a tag”). CI synchronizes its disposable package
manifest from that tag before checks and publication:

og push
og tag v0.1.0

For each subsequent release, push the intended change with og push, and
create its semantic version tag with og tag v<version> (for example, og tag v0.2.0-beta.1). Tags must be v<semver> (for example v0.1.0 or
v0.2.0-beta.1). Every purely numeric prerelease segment must be 0 or a
non-zero number without leading zeroes: v1.2.3-0 and v1.2.3-alpha01 are
valid, while v1.2.3-01, v1.2.3-0.01, and v1.2.3-alpha.01 are rejected.
Build metadata such as v1.2.3+build.1 is accepted. Stable tags publish to
npm as latest; prerelease tags publish as beta. The verify job performs an
immutable install, typecheck, tests, build, packed-artifact validation, and the
disposable DSH package smoke check before uploading the tarball consumed by the
publish job. Publishing uses npm OIDC provenance in the protected npm
environment with id-token: write; no npm token or repository secret is
configured or required for automated releases. The one-time bootstrap
publication uses the maintainer's local npm authentication only.

Development

bun install --frozen-lockfile
bun run typecheck
bun run test
bun run build
GITHUB_REF_NAME=v0.1.0 bun run release:check
bun run pack-smoke

The smoke test uses test-owned temporary directories and never modifies a live
DSH profile, credential, or production service.