🎙️ dsh-talk

dsh-talk closes the voice loop in both directions:

• speak tool — the agent speaks its replies aloud. TTS engines: the browser voice, edge-tts (network neural voices), or piper (local). Audio plays in the browser; on hosts that can carry it, the session log records the sanitized utterance (see Security boundaries).
• Composer mic button — press it, speak, and the transcription lands in the input box (or submits directly). STT engines: the browser's Web Speech (interim results included), a FunASR HTTP server, or local whisper.cpp.
• Speak-to-interrupt — starting to talk stops whatever is playing (client → host over the talk Remote namespace).
• Event announcements — turn completion, pending approvals (waterfall-safe: never blocks the gate), and errors, with a mute switch and configurable phrases.
• Settings tab — engine/language selects and announcement switches, saved as append-only profile-patch operations with backups.

browser                                host
  🎙 press ──▶ interrupt ─────────────────▶ talk/interrupt
  record (MediaRecorder / Web Speech)
  transcribe (browser) or talk/transcribe ─▶ FunASR / whisper.cpp
  setDraft(text) or submit()  ◀── talk:speech projection ── speak tool / announcements
  ▶ play audio (talk/audio or speechSynthesis)

# 1. install the bundle into your profile
dsh plugin --profile web add "github:PerryLink/dsh-talk#main"

# or from npm (published releases)
dsh plugin --profile web add dsh-talk

# 2. restart and verify the row
dsh --profile web --dump-config | grep -A2 'id: talk'

Then press the microphone next to the composer and talk; ask the agent to speak its reply:

> Say "hello" with the speak tool.

• git channel (latest main): dsh plugin --profile web add "github:PerryLink/dsh-talk#main" — the prepare script builds with production dependencies only.
• npm channel (published releases): dsh plugin --profile web add dsh-talk.
• tarball channel: pnpm pack in this repo, then dsh plugin --profile web add ./dsh-talk-<version>.tgz.
• uninstall: dsh plugin --profile web remove dsh-talk (or remove the row from the profile patch).

If pnpm reports ERR_PNPM_IGNORED_BUILDS for this package (esbuild's harmless platform-binary validation), add allowBuilds: { esbuild: true } to your pnpm-workspace.yaml — the dsh CLI prints the exact snippet.

All tunables are Schemastery Config fields (changeable from cordis.yml). cordis.patch.yml documents each key inline.

stt.silenceFinaliseMs and record.vad.silenceMs are separate mechanisms: the first finalises the Web Speech transcript when continuous recognition hears no speech, the second is the MediaRecorder energy-based detector that ends the recording (and submits it when record.autoSubmit is on). They run in different pipelines and share no state.

• Permissions: the plugin stores nothing but an in-memory, byte-capped audio cache; microphone permission is browser-mediated. The settings tab only appends patch fragments to the profile with a timestamped backup — never rewrites the file.
• Data: audio never enters the model context or the session log. Where the host's session vocabulary accepts it, the dsh-talk/speech event carries the utterance id, engine, reason, size, sanitized text, and browser voice/rate/pitch when applicable; on envelope-less hosts the event is not written at all. All display/log surfaces redact credentials, JWTs, bearer headers, and temp paths.
• Network: only the engines you configure are contacted. edge-tts performs network synthesis, FunASR uses its configured endpoint, and Chrome's webkitSpeechRecognition sends microphone audio to Google's servers for transcription; browser speechSynthesis playback remains local.

• Model-visible ⟺ logged — the model sees only the speak tool's canonical value and render text. The dsh-talk/speech event is appended only when the host can carry it (see Host compatibility); the tool/call + tool/result events always remain the reconstructable trail.
• Approval announcements never block — the approval/request listener always calls next().
• Sanitized output — credentials and temp audio paths never reach logs or displays.
• Host compatibility — the dsh-talk/speech event is appended through an adaptive gate. Hosts whose known-type vocabulary covers the event append it plainly; hosts with the ignorable append option append it with the marker; envelope-less hosts — every released line through 0.1.1-rc.2, and 0.1.2-alpha.1, which removed the envelope and fails closed on unknown types — get no append, so speech can never pollute the session log on those lines. There, the live playback history stays empty and the speak tool results are the reconstructable audit trail.
• Fail loud — invalid engines, out-of-range values, and engines configured without their required model/endpoint fail the mount.

• Browser support: Web Speech and MediaRecorder are feature-detected; without them the mic button disables itself and host engines (FunASR/whisper.cpp) can still transcribe when configured.
• Local engines are your install: edge-tts, piper, and whisper.cpp executables and models must be installed separately.
• Recording format: the browser records with its native MediaRecorder codec; whisper.cpp may require a WAV-configured recorder or a server-side conversion for other formats.
• Settings apply on reload: the settings tab appends to the profile patch; a profile reload (or web-app restart) activates the changes.
• Live playback history is empty on envelope-less hosts: on 0.1.1-rc.2 and 0.1.2-alpha.1 the host vocabulary does not know dsh-talk/speech, so the gate writes nothing and the client's session-scoped playback list stays empty. Speech itself, the mic, the settings tab, and the tool are unaffected.
• Legacy logs written by dsh-talk ≤ 0.2.1 may need repair before cold load: versions through 0.2.1 appended unmarked dsh-talk/speech events. On hosts 0.1.0-rc.7 and newer, a session whose log already contains them fails its next cold load with SessionFormatUnsupportedError. Repair: stop the host, back up the session's .jsonl log, add "ignorable":true as a top-level member of every JSON line whose "type" is "dsh-talk/speech" (for example, insert "ignorable":true, right after the opening {), then reopen the session. Nothing else changes and nothing is lost; new appends from this version never add unmarked events.

pnpm install        # node ^22.19 || >=24
pnpm run typecheck  # tsc: src + tests against the local harness checkout
pnpm run typecheck:ci  # tsc against the published 0.1.1-rc.2 types (no paths)
pnpm test           # vitest: 77 tests, 13 suites
pnpm run build      # tsc declarations + tsdown bundles (lib/)
pnpm run verify:self-contained  # dependency specs resolve from the registry
pnpm run verify:artifacts       # built ESM faces + client ModuleLoader handshake
pnpm pack           # the published tarball

dsh, dsh-plugin, deepseek-harness, deepseek, cordis, voice, speech, tts, stt, speech-to-text, text-to-speech, microphone

• @PerryLink — creator and maintainer: speak pipeline, speech engines, mic recorder, event announcements, projection unit, and the five-language docs.

This project is one of the 33 DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:

Apache License 2.0 © 2026 dsh-talk contributors