[Speech Pipeline] provider/voice/model changed mid-session,

Error message

[Speech Pipeline] provider/voice/model changed mid-session, tearing down

What it means

A watcher on [activeSpeechProvider, activeSpeechVoice.id, activeSpeechModel] cancels the current speech session whenever any of the three changes mid-session. The session is torn down via currentSession.cancel('provider-or-voice-changed') and in-flight LLM tokens for that turn are silently dropped rather than fork-replayed into the new provider (whose voice/model may differ).

Source

Thrown at packages/stage-ui/src/components/scenes/Stage.vue:899

  // })

  // await db.value?.execute(`INSERT INTO memory_test (vec) VALUES (${JSON.stringify(res.embedding)});`)
}))

// Mid-session provider / voice / model swaps would otherwise keep feeding
// tokens to the OLD adapter (segmenter for the new provider, or stale ws
// for the streaming provider). Cancel the active session so the next LLM
// token after the swap falls through `currentSession?.` cleanly (silent
// drop is acceptable — we don't try to fork-replay text into a new
// adapter with potentially different voice/model).
watch(
  [activeSpeechProvider, () => activeSpeechVoice.value?.id, activeSpeechModel],
  ([provider, voiceId, model], [prevProvider, prevVoiceId, prevModel]) => {
    if (!currentSession)
      return
    if (provider === prevProvider && voiceId === prevVoiceId && model === prevModel)
      return
    console.warn('[Speech Pipeline] provider/voice/model changed mid-session, tearing down', {
      provider,
      prevProvider,
      voiceId,
      prevVoiceId,
      model,
      prevModel,
    })
    currentSession.cancel('provider-or-voice-changed')
    currentSession = null
  },
)

// Resume audio context on first user interaction (browser requirement)
let audioContextResumed = false
function resumeAudioContextOnInteraction() {
  if (audioContextResumed || !audioContext)
    return
  audioContextResumed = true

View on GitHub (pinned to 677329427f)

Solutions

Avoid mutating speech provider/voice/model while a reply is being voiced; apply changes between turns
If teardown fires right after startup with no user action, make settings hydration write identical values without churn so the watcher does not trigger spuriously
Re-send the message to hear it with the new provider/voice
For smoother UX, defer applying speech settings until the current session finishes

Defensive patterns

Strategy: fallback

Validate before calling

// Freeze speech identity while a session is active
const speechIdentityFrozen = computed(() => currentSession !== null)
// In settings UI: disable voice/model/provider pickers while speechIdentityFrozen

Type guard

function speechIdentityChanged(
  cur: [string, string | undefined, string],
  prev: [string, string | undefined, string],
): boolean {
  return cur[0] !== prev[0] || cur[1] !== prev[1] || cur[2] !== prev[2]
}

Prevention

Defer speech settings changes until the current session finishes instead of cancelling mid-turn
Ensure settings hydration writes identical values without object/string identity churn so the watcher does not fire spuriously
After a teardown, re-send the message if the user needs to hear it with the new settings

When it happens

Trigger: Any change to speech provider, selected voice id, or speech model while a TTS session is active: the user edits speech settings mid-conversation, or a store hydration/sync rewrites these refs after speech already started.

Common situations: Changing voice while the character is speaking; switching provider mid-turn; settings hydration firing after session open, producing a spurious teardown with identical values replaced by new references.

Related errors

AI-assisted analysis of moeru-ai/airi@677329427f (2026-08-18). Data as JSON: /api/errors/e0c0da86d6d3f66b. Report an issue: GitHub.