From 66d7ef207e8affa14f338ed89b170e989e6b31d6 Mon Sep 17 00:00:00 2001
From: Neko Ayaka <neko@ayaka.moe>
Date: Sun, 9 Aug 2026 18:22:29 +0800
Subject: [PATCH] refactor(stage-web,stage-pages,stage-ui): better structure
 for asr & vad

---
 apps/stage-web/src/pages/index.vue            |  40 +-
 ...ders-transcription-realtime-aliyun-nls.vue |   9 +-
 .../aliyun-nls-transcription.vue              |   2 +-
 .../transcription/browser-web-speech-api.vue  |   2 +-
 packages/stage-shared/src/perf/io-trace.ts    |   4 +
 packages/stage-ui/package.json                |   4 +-
 .../libs/audio/vad-streaming-session.test.ts  |  50 ++
 .../src/libs/audio/vad-streaming-session.ts   | 107 ++++
 packages/stage-ui/src/libs/audio/vad.ts       |   2 +
 .../providers/providers/aliyun-nls/index.ts   |   8 +-
 .../providers/aliyun-nls/provider.ts}         | 227 ++-------
 .../providers/aliyun-nls/session.ts}          |   0
 .../providers/aliyun-nls}/token.test.ts       |   0
 .../providers/providers/aliyun-nls}/token.ts  |   0
 .../providers/providers/aliyun-nls}/utils.ts  |   0
 .../providers/browser-web-speech-api/index.ts |   5 +-
 .../browser-web-speech-api/provider.ts}       |   0
 .../providers/providers/elevenlabs/index.ts   |   2 +-
 .../providers/elevenlabs/list-models.ts       |   0
 .../stream-transcription/index.test.ts        |  34 ++
 .../providers/stream-transcription/index.ts   | 153 ++++++
 .../src/libs/providers/validators/run.test.ts |  46 ++
 .../src/libs/providers/validators/run.ts      |  39 +-
 .../stage-ui/src/stores/ai/models/vad.test.ts |  93 +++-
 packages/stage-ui/src/stores/ai/models/vad.ts |  27 +-
 .../stage-ui/src/stores/modules/hearing.ts    | 464 ++++++++++--------
 packages/stage-ui/src/workers/vad/vad.ts      |  21 +
 .../audio-transcription-stream/session.ts     |   2 +-
 28 files changed, 913 insertions(+), 428 deletions(-)
 create mode 100644 packages/stage-ui/src/libs/audio/vad-streaming-session.test.ts
 create mode 100644 packages/stage-ui/src/libs/audio/vad-streaming-session.ts
 rename packages/stage-ui/src/{stores/providers/aliyun/stream-transcription.ts => libs/providers/providers/aliyun-nls/provider.ts} (62%)
 rename packages/stage-ui/src/{stores/providers/aliyun/index.ts => libs/providers/providers/aliyun-nls/session.ts} (100%)
 rename packages/stage-ui/src/{stores/providers/aliyun => libs/providers/providers/aliyun-nls}/token.test.ts (100%)
 rename packages/stage-ui/src/{stores/providers/aliyun => libs/providers/providers/aliyun-nls}/token.ts (100%)
 rename packages/stage-ui/src/{stores/providers/aliyun => libs/providers/providers/aliyun-nls}/utils.ts (100%)
 rename packages/stage-ui/src/{stores/providers/web-speech-api/index.ts => libs/providers/providers/browser-web-speech-api/provider.ts} (100%)
 rename packages/stage-ui/src/{stores => libs/providers}/providers/elevenlabs/list-models.ts (100%)
 create mode 100644 packages/stage-ui/src/libs/providers/stream-transcription/index.test.ts
 create mode 100644 packages/stage-ui/src/libs/providers/stream-transcription/index.ts
 create mode 100644 packages/stage-ui/src/libs/providers/validators/run.test.ts

diff --git a/apps/stage-web/src/pages/index.vue b/apps/stage-web/src/pages/index.vue
index 9668fbe02..1fa2221fc 100644
--- a/apps/stage-web/src/pages/index.vue
+++ b/apps/stage-web/src/pages/index.vue
@@ -45,7 +45,7 @@ const settingsAudioDeviceStore = useSettingsAudioDevice()
 const { stream, enabled } = storeToRefs(settingsAudioDeviceStore)
 const { startRecord, stopRecord, onStopRecord } = useAudioRecorder(stream)
 const hearingPipeline = useHearingSpeechInputPipeline()
-const { transcribeForRecording } = hearingPipeline
+const { stopStreamingTranscription, transcribeForMediaStream, transcribeForRecording } = hearingPipeline
 const { supportsStreamInput } = storeToRefs(hearingPipeline)
 const providersStore = useProviderStore()
 const consciousnessStore = useConsciousnessStore()
@@ -67,28 +67,39 @@ const {
 
 let stopOnStopRecord: (() => void) | undefined
 
+async function sendVoiceInputTextToChat(text: string | undefined) {
+  if (!text?.trim())
+    return
+
+  try {
+    const provider = await providersStore.getProviderInstance(activeChatProvider.value)
+    if (!provider || !activeChatModel.value)
+      return
+
+    await chatStore.ingest(text, { model: activeChatModel.value, chatProvider: provider as ChatProvider })
+  }
+  catch (error) {
+    console.error('Failed to send chat from voice:', error)
+  }
+}
+
 async function startAudioInteraction() {
   try {
     await initVAD()
     if (stream.value)
       await startVAD(stream.value)
 
+    if (shouldUseStreamInput.value && stream.value) {
+      await transcribeForMediaStream(stream.value, {
+        onSentenceEnd: text => void sendVoiceInputTextToChat(text),
+      })
+      return
+    }
+
     // Hook once
     stopOnStopRecord = onStopRecord(async (recording) => {
       const text = await transcribeForRecording(recording)
-      if (!text || !text.trim())
-        return
-
-      try {
-        const provider = await providersStore.getProviderInstance(activeChatProvider.value)
-        if (!provider || !activeChatModel.value)
-          return
-
-        await chatStore.ingest(text, { model: activeChatModel.value, chatProvider: provider as ChatProvider })
-      }
-      catch (err) {
-        console.error('Failed to send chat from voice:', err)
-      }
+      await sendVoiceInputTextToChat(text)
     })
   }
   catch (e) {
@@ -119,6 +130,7 @@ function stopAudioInteraction() {
   try {
     stopOnStopRecord?.()
     stopOnStopRecord = undefined
+    void stopStreamingTranscription(true)
     disposeVAD()
   }
   catch {}
diff --git a/packages/stage-pages/src/pages/devtools/providers-transcription-realtime-aliyun-nls.vue b/packages/stage-pages/src/pages/devtools/providers-transcription-realtime-aliyun-nls.vue
index bc19e0255..b7c159c71 100644
--- a/packages/stage-pages/src/pages/devtools/providers-transcription-realtime-aliyun-nls.vue
+++ b/packages/stage-pages/src/pages/devtools/providers-transcription-realtime-aliyun-nls.vue
@@ -1,10 +1,11 @@
 <script setup lang="ts">
-import type { ServerEvent, ServerEvents } from '@proj-airi/stage-ui/stores/providers/aliyun'
+import type { ServerEvent, ServerEvents } from '@proj-airi/stage-ui/libs/providers/providers/aliyun-nls'
 
 import vadWorkletUrl from '@proj-airi/stage-ui/workers/vad/process.worklet?worker&url'
 
 import { errorMessageFromValue } from '@proj-airi/stage-shared'
-import { createAliyunNLSProvider, streamAliyunTranscription } from '@proj-airi/stage-ui/stores/providers/aliyun/stream-transcription'
+import { createAliyunNLSProvider } from '@proj-airi/stage-ui/libs/providers/providers/aliyun-nls'
+import { streamTranscription } from '@proj-airi/stage-ui/libs/providers/stream-transcription'
 import { Button, FieldCombobox, FieldInput } from '@proj-airi/ui'
 import { computed, nextTick, onBeforeUnmount, reactive, ref, shallowRef, watch } from 'vue'
 
@@ -159,7 +160,7 @@ async function startRecording() {
 
   appendLog('Initializing realtime transcription session')
 
-  const transcriptionResult = streamAliyunTranscription({
+  const transcriptionResult = streamTranscription({
     ...createAliyunNLSProvider(
       credentials.accessKeyId.trim(),
       credentials.accessKeySecret.trim(),
@@ -184,7 +185,7 @@ async function startRecording() {
       },
     }),
     inputAudioStream: audioStream,
-  } as unknown as Parameters<typeof streamAliyunTranscription>[0])
+  })
   transcriptionTextPromise.value = transcriptionResult.text
   isTranscribing.value = true
 
diff --git a/packages/stage-pages/src/pages/settings/providers/transcription/aliyun-nls-transcription.vue b/packages/stage-pages/src/pages/settings/providers/transcription/aliyun-nls-transcription.vue
index 911567f4f..82d026582 100644
--- a/packages/stage-pages/src/pages/settings/providers/transcription/aliyun-nls-transcription.vue
+++ b/packages/stage-pages/src/pages/settings/providers/transcription/aliyun-nls-transcription.vue
@@ -1,6 +1,6 @@
 <script setup lang="ts">
+import type { ServerEvent, ServerEvents } from '@proj-airi/stage-ui/libs/providers/providers/aliyun-nls'
 import type { HearingTranscriptionResult } from '@proj-airi/stage-ui/stores/modules/hearing'
-import type { ServerEvent, ServerEvents } from '@proj-airi/stage-ui/stores/providers/aliyun'
 import type { RemovableRef } from '@vueuse/core'
 import type { TranscriptionProviderWithExtraOptions } from '@xsai-ext/providers/utils'
 
diff --git a/packages/stage-pages/src/pages/settings/providers/transcription/browser-web-speech-api.vue b/packages/stage-pages/src/pages/settings/providers/transcription/browser-web-speech-api.vue
index b83b6e083..d620a4a95 100644
--- a/packages/stage-pages/src/pages/settings/providers/transcription/browser-web-speech-api.vue
+++ b/packages/stage-pages/src/pages/settings/providers/transcription/browser-web-speech-api.vue
@@ -10,9 +10,9 @@ import {
   ProviderSettingsLayout,
 } from '@proj-airi/stage-ui/components'
 import { selectProviderMetadata } from '@proj-airi/stage-ui/libs'
+import { streamWebSpeechAPITranscription } from '@proj-airi/stage-ui/libs/providers/providers/browser-web-speech-api'
 import { useProviderConfigStore } from '@proj-airi/stage-ui/stores/providers/config'
 import { useProviderStore } from '@proj-airi/stage-ui/stores/providers/provider'
-import { streamWebSpeechAPITranscription } from '@proj-airi/stage-ui/stores/providers/web-speech-api'
 import { useSettingsAudioDevice } from '@proj-airi/stage-ui/stores/settings'
 import { Button, FieldCombobox } from '@proj-airi/ui'
 import { until } from '@vueuse/core'
diff --git a/packages/stage-shared/src/perf/io-trace.ts b/packages/stage-shared/src/perf/io-trace.ts
index 0781c46e7..d14bf4efe 100644
--- a/packages/stage-shared/src/perf/io-trace.ts
+++ b/packages/stage-shared/src/perf/io-trace.ts
@@ -1,4 +1,5 @@
 export const IOSubsystems = {
+  VAD: 'vad',
   ASR: 'asr',
   LLM: 'llm',
   StreamingControl: 'streaming-control',
@@ -9,6 +10,7 @@ export type IOSubsystem = (typeof IOSubsystems)[keyof typeof IOSubsystems]
 
 export const IOSpanNames = {
   InteractionTurn: 'Interaction turn',
+  VoiceActivityDetection: 'Voice activity detection',
   SpeechRecognition: 'Speech recognition',
   LLMInference: 'LLM inference',
   StreamingControlDispatch: 'Streaming control dispatch',
@@ -26,6 +28,8 @@ export const IOAttributes = {
   Subsystem: `${customPrefix}.subsystem`,
   TooltipKeys: `${customPrefix}.tooltip.keys`,
   LLM_TTFT: `${customPrefix}.llm.time_to_first_token`,
+  VADAudioDurationMs: `${customPrefix}.vad.audio_duration_ms`,
+  VADAborted: `${customPrefix}.vad.aborted`,
   ASRText: `${customPrefix}.asr.text`,
   ASRAbort: `${customPrefix}.asr.abort`,
   LLMTextLength: `${customPrefix}.llm.text_length`,
diff --git a/packages/stage-ui/package.json b/packages/stage-ui/package.json
index 92f032783..96de34bf7 100644
--- a/packages/stage-ui/package.json
+++ b/packages/stage-ui/package.json
@@ -32,11 +32,13 @@
     "./libs/analytics": "./src/libs/analytics/index.ts",
     "./libs/analytics/*": "./src/libs/analytics/*.ts",
     "./libs/pinia": "./src/libs/pinia/index.ts",
+    "./libs/providers/stream-transcription": "./src/libs/providers/stream-transcription/index.ts",
+    "./libs/providers/providers/aliyun-nls": "./src/libs/providers/providers/aliyun-nls/index.ts",
+    "./libs/providers/providers/browser-web-speech-api": "./src/libs/providers/providers/browser-web-speech-api/index.ts",
     "./libs/*": "./src/libs/*.ts",
     "./libs": "./src/libs/index.ts",
     "./services/*": "./src/services/*.ts",
     "./tools/mcp": "./src/tools/mcp.ts",
-    "./stores/providers/aliyun": "./src/stores/providers/aliyun/index.ts",
     "./stores/character": "./src/stores/character/index.ts",
     "./stores/settings/analytics": "./src/stores/settings/analytics.ts",
     "./stores/settings": "./src/stores/settings/index.ts",
diff --git a/packages/stage-ui/src/libs/audio/vad-streaming-session.test.ts b/packages/stage-ui/src/libs/audio/vad-streaming-session.test.ts
new file mode 100644
index 000000000..9e29d0aec
--- /dev/null
+++ b/packages/stage-ui/src/libs/audio/vad-streaming-session.test.ts
@@ -0,0 +1,50 @@
+import { describe, expect, it, vi } from 'vitest'
+
+import { createVadStreamingSession } from './vad-streaming-session'
+
+describe('createVadStreamingSession', () => {
+  it('starts one transcription session for a detected speech segment and stops it after silence', async () => {
+    const start = vi.fn(async () => {})
+    const stop = vi.fn(async () => {})
+    const session = createVadStreamingSession({ start, stop })
+
+    session.onSpeechStart()
+    session.onSpeechStart()
+    session.onSpeechEnd()
+
+    await vi.waitFor(() => expect(stop).toHaveBeenCalledTimes(1))
+
+    expect(start).toHaveBeenCalledTimes(1)
+    expect(stop).toHaveBeenCalledTimes(1)
+  })
+
+  it('stops a session when speech ends before its asynchronous start completes', async () => {
+    let releaseStart!: () => void
+    const start = vi.fn(async () => await new Promise<void>((resolve) => {
+      releaseStart = resolve
+    }))
+    const stop = vi.fn(async () => {})
+    const session = createVadStreamingSession({ start, stop })
+
+    session.onSpeechStart()
+    await vi.waitFor(() => expect(start).toHaveBeenCalledTimes(1))
+    session.onSpeechEnd()
+    releaseStart()
+
+    await vi.waitFor(() => expect(stop).toHaveBeenCalledTimes(1))
+
+    expect(stop).toHaveBeenCalledTimes(1)
+  })
+
+  it('does not start another session after disposal', async () => {
+    const start = vi.fn(async () => {})
+    const stop = vi.fn(async () => {})
+    const session = createVadStreamingSession({ start, stop })
+
+    await session.dispose()
+    session.onSpeechStart()
+
+    expect(start).not.toHaveBeenCalled()
+    expect(stop).not.toHaveBeenCalled()
+  })
+})
diff --git a/packages/stage-ui/src/libs/audio/vad-streaming-session.ts b/packages/stage-ui/src/libs/audio/vad-streaming-session.ts
new file mode 100644
index 000000000..b5fb6b211
--- /dev/null
+++ b/packages/stage-ui/src/libs/audio/vad-streaming-session.ts
@@ -0,0 +1,107 @@
+export interface VadStreamingSessionOptions {
+  start: () => Promise<void>
+  stop: () => Promise<void>
+  onError?: (error: unknown) => void
+}
+
+/**
+ * Serializes realtime transcription sessions from VAD speech boundaries.
+ *
+ * A detected speech segment owns one provider session. The session starts when
+ * VAD detects speech and stops after VAD reports the configured silence period.
+ */
+export function createVadStreamingSession(options: VadStreamingSessionOptions) {
+  let disposed = false
+  let speechActive = false
+  let providerSessionActive = false
+  let lifecycle = Promise.resolve()
+
+  function enqueue(operation: () => Promise<void>) {
+    lifecycle = lifecycle
+      .catch(() => undefined)
+      .then(operation)
+    return lifecycle
+  }
+
+  function onSpeechStart() {
+    if (disposed || speechActive)
+      return
+
+    speechActive = true
+    void enqueue(async () => {
+      if (disposed || providerSessionActive)
+        return
+
+      try {
+        await options.start()
+        providerSessionActive = true
+      }
+      catch (error) {
+        options.onError?.(error)
+        return
+      }
+
+      if (!disposed && speechActive)
+        return
+
+      try {
+        await options.stop()
+      }
+      catch (error) {
+        options.onError?.(error)
+      }
+      finally {
+        providerSessionActive = false
+      }
+    })
+  }
+
+  function onSpeechEnd() {
+    if (disposed || !speechActive)
+      return
+
+    speechActive = false
+    void enqueue(async () => {
+      if (!providerSessionActive)
+        return
+
+      try {
+        await options.stop()
+      }
+      catch (error) {
+        options.onError?.(error)
+      }
+      finally {
+        providerSessionActive = false
+      }
+    })
+  }
+
+  async function dispose() {
+    if (disposed)
+      return
+
+    disposed = true
+    speechActive = false
+    await enqueue(async () => {
+      if (!providerSessionActive)
+        return
+
+      try {
+        await options.stop()
+      }
+      catch (error) {
+        options.onError?.(error)
+      }
+      finally {
+        providerSessionActive = false
+      }
+    })
+  }
+
+  return {
+    onSpeechStart,
+    onSpeechEnd,
+    dispose,
+  }
+}
diff --git a/packages/stage-ui/src/libs/audio/vad.ts b/packages/stage-ui/src/libs/audio/vad.ts
index f9bbbc2f2..4a50d9e4b 100644
--- a/packages/stage-ui/src/libs/audio/vad.ts
+++ b/packages/stage-ui/src/libs/audio/vad.ts
@@ -20,6 +20,8 @@ export interface BaseVADConfig {
 export interface VADEvents {
   // Emitted when speech is detected
   'speech-start': void
+  // Emitted with each PCM chunk that belongs to the active speech segment
+  'speech-audio': { buffer: Float32Array }
   // Emitted when speech has ended
   'speech-end': void
   // Emitted when a complete speech segment is ready for transcription
diff --git a/packages/stage-ui/src/libs/providers/providers/aliyun-nls/index.ts b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/index.ts
index 687269f83..9fc7e7d21 100644
--- a/packages/stage-ui/src/libs/providers/providers/aliyun-nls/index.ts
+++ b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/index.ts
@@ -1,11 +1,11 @@
 import type { TranscriptionProviderWithExtraOptions } from '@xsai-ext/providers/utils'
 
-import type { AliyunRealtimeSpeechExtraOptions } from '../../../../stores/providers/aliyun/stream-transcription'
+import type { AliyunRealtimeSpeechExtraOptions } from './provider'
 
 import { z } from 'zod'
 
-import { createAliyunNLSProvider } from '../../../../stores/providers/aliyun/stream-transcription'
 import { defineProvider } from '../registry'
+import { createAliyunNLSProvider } from './provider'
 
 const aliyunNlsRegions = [
   'cn-shanghai',
@@ -105,3 +105,7 @@ export const providerAliyunNlsTranscription = defineProvider<AliyunNlsConfig>({
     }],
   },
 })
+
+export type { AliyunRealtimeSpeechExtraOptions } from './provider'
+export { createAliyunNLSProvider } from './provider'
+export type { ServerEvent, ServerEvents } from './session'
diff --git a/packages/stage-ui/src/stores/providers/aliyun/stream-transcription.ts b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/provider.ts
similarity index 62%
rename from packages/stage-ui/src/stores/providers/aliyun/stream-transcription.ts
rename to packages/stage-ui/src/libs/providers/providers/aliyun-nls/provider.ts
index 3e81288f7..21e27b153 100644
--- a/packages/stage-ui/src/stores/providers/aliyun/stream-transcription.ts
+++ b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/provider.ts
@@ -1,13 +1,11 @@
 import type { SpeechProviderWithExtraOptions } from '@xsai-ext/providers/utils'
-import type { CommonRequestOptions } from '@xsai/shared'
-import type { StreamTranscriptionDelta, StreamTranscriptionResult } from '@xsai/stream-transcription'
 
-import type { EventStartTranscription, ServerEvent, ServerEvents } from './'
+import type { EventStartTranscription, ServerEvent, ServerEvents } from './session'
 
 import { tryCatch } from '@moeru/std'
 import { timeout as promiseTimeout } from 'es-toolkit/promise'
 
-import { createAliyunNLSSession } from './'
+import { createAliyunNLSSession } from './session'
 import { nlsWebSocketEndpointFromRegion } from './utils'
 
 type SessionOptions = NonNullable<Parameters<typeof createAliyunNLSSession>[3]>
@@ -90,15 +88,6 @@ export interface AliyunStreamTranscriptionHandle {
   close: () => Promise<void>
 }
 
-interface AliyunStreamTranscriptionOptions extends AliyunRealtimeSpeechExtraOptions {
-  baseURL?: CommonRequestOptions['baseURL']
-  fetch?: CommonRequestOptions['fetch']
-  headers?: HeadersInit
-  file?: Blob
-  fileName?: string
-  inputStream?: ReadableStream<AudioChunk>
-}
-
 function toArrayBuffer(chunk: AudioChunk): ArrayBuffer {
   if (chunk instanceof ArrayBuffer)
     return chunk
@@ -115,71 +104,10 @@ function toArrayBuffer(chunk: AudioChunk): ArrayBuffer {
 
 const sseEncoder = new TextEncoder()
 
-function encodeSSE(payload: StreamTranscriptionDelta): Uint8Array {
+function encodeSSE(payload: { delta: string, type: 'transcript.text.delta' | 'transcript.text.done' }): Uint8Array {
   return sseEncoder.encode(`data: ${JSON.stringify(payload)}\n\n`)
 }
 
-// NOTICE: Copied/adapted from @xsai/stream-transcription SSE parsing to keep behavior consistent.
-// Ref: @xsai/stream-transcription@0.4.0-beta.8 (dist/index.js parseChunk/transformChunk).
-function parseSSELine(line: string): StreamTranscriptionDelta | undefined {
-  if (!line || !line.startsWith('data:'))
-    return undefined
-
-  const content = line.slice('data:'.length)
-  const data = content.startsWith(' ') ? content.slice(1) : content
-  if (!data)
-    return undefined
-
-  return JSON.parse(data) as StreamTranscriptionDelta
-}
-
-function aliyunChunkTransformer() {
-  const decoder = new TextDecoder()
-  let buffer = ''
-
-  return new TransformStream<Uint8Array, StreamTranscriptionDelta>({
-    transform: (chunk, controller) => {
-      buffer += decoder.decode(chunk, { stream: true })
-      const lines = buffer.split('\n')
-      buffer = lines.pop() ?? ''
-
-      for (const line of lines) {
-        const parsed = parseSSELine(line)
-        if (parsed)
-          controller.enqueue(parsed)
-      }
-    },
-    flush: (controller) => {
-      if (!buffer)
-        return
-      const parsed = parseSSELine(buffer)
-      if (parsed)
-        controller.enqueue(parsed)
-    },
-  })
-}
-
-// NOTICE: Copied/adapted from @xsai/stream-transcription delayed promise helper.
-// Ref: @xsai/stream-transcription@0.4.0-beta.8 (dist/index.js DelayedPromise usage).
-function createDeferred<T>() {
-  let resolve!: (value: T | PromiseLike<T>) => void
-  let reject!: (reason?: unknown) => void
-  const promise = new Promise<T>((res, rej) => {
-    resolve = res
-    reject = rej
-  })
-
-  return { promise, resolve, reject }
-}
-
-function resolveAudioStream(options: AliyunStreamTranscriptionOptions): ReadableStream<AudioChunk> {
-  const stream = options.inputAudioStream ?? options.inputStream ?? options.file?.stream()
-  if (!stream)
-    throw new TypeError('Audio stream or file is required for Aliyun streaming transcription.')
-
-  return stream as ReadableStream<AudioChunk>
-}
-
 interface InternalRealtimeOptions extends CreateAliyunStreamTranscriptionOptions {
   onSentenceFinal?: (payload: ServerEvents['SentenceEnd']) => Promise<void> | void
   idleTimeoutMs?: number
@@ -219,6 +147,7 @@ async function startRealtimeSession(options: InternalRealtimeOptions): Promise<A
 
   const stopWaiter = createWaiter(stopAckTimeoutMs, abortSignal)
   let stopping = false
+  let cleanupPromise: Promise<void> | undefined
 
   async function requestStop(reason?: unknown) {
     if (stopping)
@@ -252,28 +181,35 @@ async function startRealtimeSession(options: InternalRealtimeOptions): Promise<A
 
   bumpIdle()
 
-  async function cleanup(error?: unknown, options?: { sendStop?: boolean, closeSocket?: boolean }) {
-    const { sendStop = true, closeSocket = true } = options ?? {}
-    abortHandler?.off()
-    await tryCatch(async () => await reader.cancel())
+  function cleanup(error?: unknown, options?: { sendStop?: boolean, closeSocket?: boolean }) {
+    if (cleanupPromise)
+      return cleanupPromise
 
-    if (websocket && closeSocket) {
-      switch (websocket.readyState) {
-        case WebSocket.OPEN:
-          if (sendStop)
-            await tryCatch(() => session.stop(websocket))
-          websocket.close(1000, 'client closed')
-          break
-        case WebSocket.CONNECTING:
-          websocket.close(1000, 'client closed')
-          break
-        default:
-          // If the server has already initiated closure, avoid sending another close frame.
-          break
+    cleanupPromise = (async () => {
+      const { sendStop = true, closeSocket = true } = options ?? {}
+      abortHandler?.off()
+      await tryCatch(async () => await reader.cancel())
+
+      if (websocket && closeSocket) {
+        switch (websocket.readyState) {
+          case WebSocket.OPEN:
+            if (sendStop)
+              await tryCatch(() => session.stop(websocket))
+            websocket.close(1000, 'client closed')
+            break
+          case WebSocket.CONNECTING:
+            websocket.close(1000, 'client closed')
+            break
+          default:
+            // If the server has already initiated closure, avoid sending another close frame.
+            break
+        }
       }
-    }
 
-    await onSessionTerminated?.(error)
+      await onSessionTerminated?.(error)
+    })()
+
+    return cleanupPromise
   }
 
   const handle: AliyunStreamTranscriptionHandle = {
@@ -296,8 +232,9 @@ async function startRealtimeSession(options: InternalRealtimeOptions): Promise<A
         bumpIdle()
       }
 
-      // Allow a grace period for server to flush final events before stop.
-      bumpIdle()
+      // The VAD-owned audio stream closes only after speech is complete. End
+      // NLS immediately so it returns its final sentence before the next VAD segment.
+      await requestStop()
     }
     catch (error) {
       await cleanup(error)
@@ -356,87 +293,6 @@ async function startRealtimeSession(options: InternalRealtimeOptions): Promise<A
   return handle
 }
 
-export function streamAliyunTranscription(options: AliyunStreamTranscriptionOptions): StreamTranscriptionResult {
-  const audioStream = resolveAudioStream(options)
-  const fetcher = options.fetch ?? globalThis.fetch
-  const deferredText = createDeferred<string>()
-
-  let text = ''
-  let textStreamCtrl: ReadableStreamDefaultController<string> | undefined
-  let fullStreamCtrl: ReadableStreamDefaultController<StreamTranscriptionDelta> | undefined
-
-  const fullStream = new ReadableStream<StreamTranscriptionDelta>({
-    start(controller) {
-      fullStreamCtrl = controller
-    },
-  })
-
-  const textStream = new ReadableStream<string>({
-    start(controller) {
-      textStreamCtrl = controller
-    },
-  })
-
-  const doStream = async () => {
-    const requestTarget = options.baseURL instanceof URL
-      ? options.baseURL
-      : new URL(typeof options.baseURL === 'string' ? options.baseURL : 'http://localhost')
-    const response = await fetcher(requestTarget, {
-      body: audioStream,
-      headers: options.headers,
-      method: 'POST',
-      signal: options.abortSignal,
-    })
-
-    if (!response.ok)
-      throw new Error(`Aliyun streaming transcription request failed with status ${response.status}`)
-
-    if (!response.body)
-      throw new Error('Streaming transcription response is missing a readable body.')
-
-    await response.body
-      .pipeThrough(aliyunChunkTransformer())
-      .pipeTo(new WritableStream<StreamTranscriptionDelta>({
-        write: (chunk) => {
-          fullStreamCtrl?.enqueue(chunk)
-          if (chunk.type === 'transcript.text.delta') {
-            text += chunk.delta
-            textStreamCtrl?.enqueue(chunk.delta)
-          }
-        },
-        close: () => {
-          fullStreamCtrl?.close()
-          textStreamCtrl?.close()
-        },
-        abort: (reason) => {
-          fullStreamCtrl?.error(reason)
-          textStreamCtrl?.error(reason)
-        },
-      }))
-  }
-
-  void (async () => {
-    try {
-      await doStream()
-      deferredText.resolve(text)
-    }
-    catch (error) {
-      fullStreamCtrl?.error(error)
-      textStreamCtrl?.error(error)
-      deferredText.reject(error)
-    }
-  })()
-
-  // REVIEW: We mirrored the streaming orchestration from @xsai/stream-transcription instead of
-  // patching the upstream package because Aliyun uses a custom websocket+fetch bridge (no FormData).
-  // Keeping it local avoids diverging from the published package while we wait for upstream support.
-  return {
-    fullStream,
-    text: deferredText.promise,
-    textStream,
-  }
-}
-
 export function createAliyunNLSProvider(
   accessKeyId: string,
   accessKeySecret: string,
@@ -469,20 +325,21 @@ export function createAliyunNLSProvider(
                 audioStream: streamSource as ReadableStream<AudioChunk>,
                 abortSignal: extraOptions?.abortSignal || init?.signal || undefined,
                 hooks: extraOptions?.hooks,
-                onSessionTerminated: async (error) => {
+                onSessionTerminated: async (sessionError) => {
                   controllerClosed = true
                   try {
-                    await extraOptions?.onSessionTerminated?.(error)
+                    await extraOptions?.onSessionTerminated?.(sessionError)
+                    if (sessionError) {
+                      controller.error(sessionError instanceof Error ? sessionError : new Error(String(sessionError)))
+                      return
+                    }
+
                     controller.enqueue(encodeSSE({ delta: '', type: 'transcript.text.done' }))
+                    controller.close()
                   }
                   catch (error) {
                     console.error('error in onSessionTerminated hook:', error)
-                  }
-                  finally {
-                    if (error)
-                      controller.error(error instanceof Error ? error : new Error(String(error)))
-                    else
-                      controller.close()
+                    controller.error(error instanceof Error ? error : new Error(String(error)))
                   }
                 },
                 onSentenceFinal: async (payload) => {
diff --git a/packages/stage-ui/src/stores/providers/aliyun/index.ts b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/session.ts
similarity index 100%
rename from packages/stage-ui/src/stores/providers/aliyun/index.ts
rename to packages/stage-ui/src/libs/providers/providers/aliyun-nls/session.ts
diff --git a/packages/stage-ui/src/stores/providers/aliyun/token.test.ts b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/token.test.ts
similarity index 100%
rename from packages/stage-ui/src/stores/providers/aliyun/token.test.ts
rename to packages/stage-ui/src/libs/providers/providers/aliyun-nls/token.test.ts
diff --git a/packages/stage-ui/src/stores/providers/aliyun/token.ts b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/token.ts
similarity index 100%
rename from packages/stage-ui/src/stores/providers/aliyun/token.ts
rename to packages/stage-ui/src/libs/providers/providers/aliyun-nls/token.ts
diff --git a/packages/stage-ui/src/stores/providers/aliyun/utils.ts b/packages/stage-ui/src/libs/providers/providers/aliyun-nls/utils.ts
similarity index 100%
rename from packages/stage-ui/src/stores/providers/aliyun/utils.ts
rename to packages/stage-ui/src/libs/providers/providers/aliyun-nls/utils.ts
diff --git a/packages/stage-ui/src/libs/providers/providers/browser-web-speech-api/index.ts b/packages/stage-ui/src/libs/providers/providers/browser-web-speech-api/index.ts
index d87cb839e..cbd0a5554 100644
--- a/packages/stage-ui/src/libs/providers/providers/browser-web-speech-api/index.ts
+++ b/packages/stage-ui/src/libs/providers/providers/browser-web-speech-api/index.ts
@@ -1,8 +1,8 @@
 import { isStageTamagotchi } from '@proj-airi/stage-shared'
 import { z } from 'zod'
 
-import { createWebSpeechAPIProvider } from '../../../../stores/providers/web-speech-api'
 import { defineProvider } from '../registry'
+import { createWebSpeechAPIProvider } from './provider'
 
 const webSpeechApiConfigSchema = z.object({
   language: z.string().default('en-US'),
@@ -11,6 +11,9 @@ const webSpeechApiConfigSchema = z.object({
   maxAlternatives: z.number().int().positive().default(1),
 })
 
+export type { WebSpeechAPIExtraOptions } from './provider'
+export { createWebSpeechAPIProvider, streamWebSpeechAPITranscription } from './provider'
+
 function isWebSpeechApiAvailable() {
   if (typeof window === 'undefined')
     return false
diff --git a/packages/stage-ui/src/stores/providers/web-speech-api/index.ts b/packages/stage-ui/src/libs/providers/providers/browser-web-speech-api/provider.ts
similarity index 100%
rename from packages/stage-ui/src/stores/providers/web-speech-api/index.ts
rename to packages/stage-ui/src/libs/providers/providers/browser-web-speech-api/provider.ts
diff --git a/packages/stage-ui/src/libs/providers/providers/elevenlabs/index.ts b/packages/stage-ui/src/libs/providers/providers/elevenlabs/index.ts
index 928bc1c22..adec28884 100644
--- a/packages/stage-ui/src/libs/providers/providers/elevenlabs/index.ts
+++ b/packages/stage-ui/src/libs/providers/providers/elevenlabs/index.ts
@@ -5,8 +5,8 @@ import type { ComposerTranslation } from 'vue-i18n'
 import { createUnElevenLabs, listVoices } from 'unspeech'
 import { z } from 'zod'
 
-import { models as elevenLabsModels } from '../../../../stores/providers/elevenlabs/list-models'
 import { defineProvider } from '../registry'
+import { models as elevenLabsModels } from './list-models'
 
 const elevenLabsConfigSchema = z.object({
   apiKey: z.string(),
diff --git a/packages/stage-ui/src/stores/providers/elevenlabs/list-models.ts b/packages/stage-ui/src/libs/providers/providers/elevenlabs/list-models.ts
similarity index 100%
rename from packages/stage-ui/src/stores/providers/elevenlabs/list-models.ts
rename to packages/stage-ui/src/libs/providers/providers/elevenlabs/list-models.ts
diff --git a/packages/stage-ui/src/libs/providers/stream-transcription/index.test.ts b/packages/stage-ui/src/libs/providers/stream-transcription/index.test.ts
new file mode 100644
index 000000000..eaf2d2dd5
--- /dev/null
+++ b/packages/stage-ui/src/libs/providers/stream-transcription/index.test.ts
@@ -0,0 +1,34 @@
+import { describe, expect, it } from 'vitest'
+
+import { streamTranscription } from './index'
+
+describe('streamTranscription', () => {
+  it('parses split SSE chunks and joins transcription deltas', async () => {
+    const encoder = new TextEncoder()
+    const responseBody = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(encoder.encode('data: {"type":"transcript.text.delta","delta":"Hello"}\n'))
+        controller.enqueue(encoder.encode('\ndata: {"type":"transcript.text.delta","delta":" AIRI"}\n\n'))
+        controller.close()
+      },
+    })
+    const audioStream = new ReadableStream<ArrayBuffer>({
+      start(controller) {
+        controller.close()
+      },
+    })
+
+    const result = streamTranscription({
+      baseURL: 'https://example.invalid/transcription',
+      fetch: async () => new Response(responseBody),
+      inputAudioStream: audioStream,
+    })
+
+    expect(await result.text).toBe('Hello AIRI')
+    await expect(result.textStream.getReader().read()).resolves.toEqual({ done: false, value: 'Hello' })
+  })
+
+  it('rejects requests without an audio input', () => {
+    expect(() => streamTranscription({})).toThrow('Audio stream or file is required')
+  })
+})
diff --git a/packages/stage-ui/src/libs/providers/stream-transcription/index.ts b/packages/stage-ui/src/libs/providers/stream-transcription/index.ts
new file mode 100644
index 000000000..d3dd7a355
--- /dev/null
+++ b/packages/stage-ui/src/libs/providers/stream-transcription/index.ts
@@ -0,0 +1,153 @@
+import type { CommonRequestOptions } from '@xsai/shared'
+import type { StreamTranscriptionDelta, StreamTranscriptionResult } from '@xsai/stream-transcription'
+
+type AudioChunk = ArrayBuffer | ArrayBufferView
+
+/** Options for adapting an SSE transcription request to xsAI stream results. */
+export interface StreamTranscriptionOptions {
+  abortSignal?: AbortSignal
+  baseURL?: CommonRequestOptions['baseURL']
+  fetch?: CommonRequestOptions['fetch']
+  headers?: HeadersInit
+  file?: Blob
+  inputAudioStream?: ReadableStream<AudioChunk>
+  inputStream?: ReadableStream<AudioChunk>
+}
+
+function createDeferred<T>() {
+  let resolve!: (value: T | PromiseLike<T>) => void
+  let reject!: (reason?: unknown) => void
+  const promise = new Promise<T>((res, rej) => {
+    resolve = res
+    reject = rej
+  })
+
+  return { promise, resolve, reject }
+}
+
+function resolveAudioStream(options: StreamTranscriptionOptions): ReadableStream<AudioChunk> {
+  const stream = options.inputAudioStream ?? options.inputStream ?? options.file?.stream()
+  if (!stream)
+    throw new TypeError('Audio stream or file is required for streaming transcription.')
+
+  return stream as ReadableStream<AudioChunk>
+}
+
+function parseSSELine(line: string): StreamTranscriptionDelta | undefined {
+  if (!line || !line.startsWith('data:'))
+    return undefined
+
+  const content = line.slice('data:'.length)
+  const data = content.startsWith(' ') ? content.slice(1) : content
+  if (!data)
+    return undefined
+
+  return JSON.parse(data) as StreamTranscriptionDelta
+}
+
+function createSSETransformer() {
+  const decoder = new TextDecoder()
+  let buffer = ''
+
+  return new TransformStream<Uint8Array, StreamTranscriptionDelta>({
+    transform: (chunk, controller) => {
+      buffer += decoder.decode(chunk, { stream: true })
+      const lines = buffer.split('\n')
+      buffer = lines.pop() ?? ''
+
+      for (const line of lines) {
+        const parsed = parseSSELine(line)
+        if (parsed)
+          controller.enqueue(parsed)
+      }
+    },
+    flush: (controller) => {
+      if (!buffer)
+        return
+
+      const parsed = parseSSELine(buffer)
+      if (parsed)
+        controller.enqueue(parsed)
+    },
+  })
+}
+
+/**
+ * Converts an SSE transcription endpoint into xsAI's streaming result shape.
+ *
+ * The provider owns transport details. This adapter owns only request input,
+ * SSE parsing, and the result streams consumed by Hearing.
+ */
+export function streamTranscription(options: StreamTranscriptionOptions): StreamTranscriptionResult {
+  const audioStream = resolveAudioStream(options)
+  const fetcher = options.fetch ?? globalThis.fetch
+  const deferredText = createDeferred<string>()
+
+  let text = ''
+  let textStreamCtrl: ReadableStreamDefaultController<string> | undefined
+  let fullStreamCtrl: ReadableStreamDefaultController<StreamTranscriptionDelta> | undefined
+
+  const fullStream = new ReadableStream<StreamTranscriptionDelta>({
+    start(controller) {
+      fullStreamCtrl = controller
+    },
+  })
+
+  const textStream = new ReadableStream<string>({
+    start(controller) {
+      textStreamCtrl = controller
+    },
+  })
+
+  void (async () => {
+    try {
+      const requestTarget = options.baseURL instanceof URL
+        ? options.baseURL
+        : new URL(typeof options.baseURL === 'string' ? options.baseURL : 'http://localhost')
+      const response = await fetcher(requestTarget, {
+        body: audioStream,
+        headers: options.headers,
+        method: 'POST',
+        signal: options.abortSignal,
+      })
+
+      if (!response.ok)
+        throw new Error(`Streaming transcription request failed with status ${response.status}`)
+
+      if (!response.body)
+        throw new Error('Streaming transcription response is missing a readable body.')
+
+      await response.body
+        .pipeThrough(createSSETransformer())
+        .pipeTo(new WritableStream<StreamTranscriptionDelta>({
+          write: (chunk) => {
+            fullStreamCtrl?.enqueue(chunk)
+            if (chunk.type === 'transcript.text.delta') {
+              text += chunk.delta
+              textStreamCtrl?.enqueue(chunk.delta)
+            }
+          },
+          close: () => {
+            fullStreamCtrl?.close()
+            textStreamCtrl?.close()
+          },
+          abort: (reason) => {
+            fullStreamCtrl?.error(reason)
+            textStreamCtrl?.error(reason)
+          },
+        }))
+      deferredText.resolve(text)
+    }
+    catch (error) {
+      fullStreamCtrl?.error(error)
+      textStreamCtrl?.error(error)
+      deferredText.reject(error)
+    }
+  })()
+
+  return {
+    fullStream,
+    text: deferredText.promise,
+    textStream,
+  }
+}
diff --git a/packages/stage-ui/src/libs/providers/validators/run.test.ts b/packages/stage-ui/src/libs/providers/validators/run.test.ts
new file mode 100644
index 000000000..31acc2c82
--- /dev/null
+++ b/packages/stage-ui/src/libs/providers/validators/run.test.ts
@@ -0,0 +1,46 @@
+import type { ComposerTranslation } from 'vue-i18n'
+
+import type { ProviderDefinition } from '../types'
+
+import { createChatProvider } from '@xsai-ext/providers/utils'
+import { describe, expect, it, vi } from 'vitest'
+
+import { validateProvider } from './run'
+
+const mockT = ((key: string) => key) as unknown as ComposerTranslation
+
+describe('validateProvider', () => {
+  it('disposes the temporary provider after runtime validation', async () => {
+    const dispose = vi.fn()
+    const provider = Object.assign(
+      createChatProvider({ apiKey: 'test', baseURL: 'https://example.com/v1' }),
+      { dispose },
+    )
+    const definition: ProviderDefinition<Record<string, unknown>> = {
+      id: 'example',
+      name: 'Example',
+      description: 'Example provider',
+      nameLocalize: input => input.t('example'),
+      descriptionLocalize: input => input.t('example'),
+      tasks: [],
+      createProviderConfig: () => ({}) as never,
+      createProvider: () => provider,
+    }
+
+    await validateProvider({
+      steps: [{ id: 'runtime', label: 'Runtime', status: 'idle', reason: '', kind: 'provider' }],
+      config: {},
+      definition,
+      configValidators: [],
+      providerValidators: [{
+        id: 'runtime',
+        name: 'Runtime',
+        validator: async () => ({ valid: true, errors: [], reason: '', reasonKey: '' }),
+      }],
+      providerExtra: undefined,
+      shouldValidate: true,
+    }, { t: mockT })
+
+    expect(dispose).toHaveBeenCalledTimes(1)
+  })
+})
diff --git a/packages/stage-ui/src/libs/providers/validators/run.ts b/packages/stage-ui/src/libs/providers/validators/run.ts
index f084f6d44..38baad4d8 100644
--- a/packages/stage-ui/src/libs/providers/validators/run.ts
+++ b/packages/stage-ui/src/libs/providers/validators/run.ts
@@ -164,23 +164,28 @@ export async function validateProvider(
     return steps
   }
 
-  await Promise.all(providerValidators.map(async (validatorDefinition, index) => {
-    const step = steps[providerStepOffset + index]
-    step.status = 'validating'
-    step.reason = ''
-    onValidatorStart?.({ kind: 'provider', index, step })
-    try {
-      const result = await validatorDefinition.validator(config, providerInstance, providerExtra as any, runContext)
-      step.status = result.valid ? 'valid' : 'invalid'
-      step.reason = result.valid ? '' : result.reason
-      onValidatorSuccess?.({ kind: 'provider', index, step, result })
-    }
-    catch (error) {
-      step.status = 'invalid'
-      step.reason = errorMessageFrom(error) ?? 'Unknown error'
-      onValidatorError?.({ kind: 'provider', index, step, error })
-    }
-  }))
+  try {
+    await Promise.all(providerValidators.map(async (validatorDefinition, index) => {
+      const step = steps[providerStepOffset + index]
+      step.status = 'validating'
+      step.reason = ''
+      onValidatorStart?.({ kind: 'provider', index, step })
+      try {
+        const result = await validatorDefinition.validator(config, providerInstance, providerExtra as any, runContext)
+        step.status = result.valid ? 'valid' : 'invalid'
+        step.reason = result.valid ? '' : result.reason
+        onValidatorSuccess?.({ kind: 'provider', index, step, result })
+      }
+      catch (error) {
+        step.status = 'invalid'
+        step.reason = errorMessageFrom(error) ?? 'Unknown error'
+        onValidatorError?.({ kind: 'provider', index, step, error })
+      }
+    }))
+  }
+  finally {
+    await (providerInstance as ProviderInstance & { dispose?: () => Promise<void> | void }).dispose?.()
+  }
 
   return steps
 }
diff --git a/packages/stage-ui/src/stores/ai/models/vad.test.ts b/packages/stage-ui/src/stores/ai/models/vad.test.ts
index f75706483..8c7cb449f 100644
--- a/packages/stage-ui/src/stores/ai/models/vad.test.ts
+++ b/packages/stage-ui/src/stores/ai/models/vad.test.ts
@@ -1,7 +1,45 @@
-import { describe, expect, it } from 'vitest'
+import { IOAttributes, IOSpanNames, IOSubsystems } from '@proj-airi/stage-shared'
+import { beforeEach, describe, expect, it, vi } from 'vitest'
 
 import { resolveVADConfig } from './vad'
 
+const vadMocks = vi.hoisted(() => {
+  const handlers = new Map<string, (event?: unknown) => void>()
+
+  return {
+    handlers,
+    createVAD: vi.fn(async () => ({
+      on: vi.fn((name: string, handler: (event?: unknown) => void) => handlers.set(name, handler)),
+      updateConfig: vi.fn(),
+    })),
+    initialize: vi.fn(async () => undefined),
+    start: vi.fn(async () => undefined),
+    stop: vi.fn(),
+    dispose: vi.fn(),
+  }
+})
+
+const spanMock = vi.hoisted(() => ({
+  end: vi.fn(),
+  setAttribute: vi.fn(),
+}))
+
+const startSpanMock = vi.hoisted(() => vi.fn(() => spanMock))
+
+vi.mock('../../../workers/vad', () => ({
+  createVAD: vadMocks.createVAD,
+  createVADStates: () => ({
+    initialize: vadMocks.initialize,
+    start: vadMocks.start,
+    stop: vadMocks.stop,
+    dispose: vadMocks.dispose,
+  }),
+}))
+
+vi.mock('../../../composables/use-io-tracer', () => ({
+  startSpan: startSpanMock,
+}))
+
 describe('resolveVADConfig', () => {
   it('uses safer defaults for threshold and silence duration', () => {
     expect(resolveVADConfig()).toEqual({
@@ -23,3 +61,56 @@ describe('resolveVADConfig', () => {
     })
   })
 })
+
+describe('useVAD', () => {
+  beforeEach(() => {
+    vadMocks.handlers.clear()
+    vi.clearAllMocks()
+  })
+
+  it('records a completed VAD span for each detected speech segment', async () => {
+    const { useVAD } = await import('./vad')
+    const vad = useVAD('vad-worker-url')
+
+    await vad.init()
+    vadMocks.handlers.get('speech-start')?.()
+    vadMocks.handlers.get('speech-ready')?.({
+      buffer: new Float32Array([0.25, -0.25]),
+      duration: 1500,
+    })
+
+    expect(startSpanMock).toHaveBeenCalledWith(
+      IOSpanNames.VoiceActivityDetection,
+      undefined,
+      {
+        [IOAttributes.Subsystem]: IOSubsystems.VAD,
+      },
+    )
+    expect(spanMock.setAttribute).toHaveBeenCalledWith(IOAttributes.VADAudioDurationMs, 1500)
+    expect(spanMock.end).toHaveBeenCalledOnce()
+  })
+
+  it('marks an active VAD span as aborted when the session is disposed', async () => {
+    const { useVAD } = await import('./vad')
+    const vad = useVAD('vad-worker-url')
+
+    await vad.init()
+    vadMocks.handlers.get('speech-start')?.()
+    vad.dispose()
+
+    expect(spanMock.setAttribute).toHaveBeenCalledWith(IOAttributes.VADAborted, true)
+    expect(spanMock.end).toHaveBeenCalledOnce()
+  })
+
+  it('forwards VAD-owned PCM chunks while speech is active', async () => {
+    const onSpeechAudio = vi.fn()
+    const { useVAD } = await import('./vad')
+    const vad = useVAD('vad-worker-url', { onSpeechAudio })
+    const buffer = new Float32Array([0.25, -0.25])
+
+    await vad.init()
+    vadMocks.handlers.get('speech-audio')?.({ buffer })
+
+    expect(onSpeechAudio).toHaveBeenCalledWith({ buffer })
+  })
+})
diff --git a/packages/stage-ui/src/stores/ai/models/vad.ts b/packages/stage-ui/src/stores/ai/models/vad.ts
index 73d944243..47438d75e 100644
--- a/packages/stage-ui/src/stores/ai/models/vad.ts
+++ b/packages/stage-ui/src/stores/ai/models/vad.ts
@@ -1,11 +1,13 @@
+import type { Span } from '@opentelemetry/api'
 import type { MaybeRefOrGetter } from 'vue'
 
 import type { BaseVADConfig } from '../../../libs/audio/vad'
 
 import { merge } from '@moeru/std'
-import { errorMessageFromValue } from '@proj-airi/stage-shared'
+import { errorMessageFromValue, IOAttributes, IOSpanNames, IOSubsystems } from '@proj-airi/stage-shared'
 import { ref, toRef, watch } from 'vue'
 
+import { startSpan } from '../../../composables/use-io-tracer'
 import { createVAD, createVADStates } from '../../../workers/vad'
 
 interface UseVADOptions {
@@ -15,6 +17,7 @@ interface UseVADOptions {
   minSpeechDurationMs?: MaybeRefOrGetter<number>
 
   onSpeechStart?: () => void
+  onSpeechAudio?: (event: { buffer: Float32Array }) => void
   onSpeechEnd?: () => void
   onSpeechReady?: (event: { buffer: Float32Array, duration: number }) => void
 }
@@ -62,6 +65,17 @@ export function useVAD(workerUrl: string, options?: UseVADOptions) {
 
   const loaded = ref(false)
   const loading = ref(false)
+  let activeSpan: Span | undefined
+
+  function finishActiveSpan(aborted: boolean) {
+    if (!activeSpan)
+      return
+
+    if (aborted)
+      activeSpan.setAttribute(IOAttributes.VADAborted, true)
+    activeSpan.end()
+    activeSpan = undefined
+  }
 
   const threshold = toRef(options.threshold)
   const minSilenceDurationMs = toRef(options.minSilenceDurationMs)
@@ -90,16 +104,26 @@ export function useVAD(workerUrl: string, options?: UseVADOptions) {
 
       // Set up event handlers
       vad.value.on('speech-start', () => {
+        finishActiveSpan(true)
+        activeSpan = startSpan(IOSpanNames.VoiceActivityDetection, undefined, {
+          [IOAttributes.Subsystem]: IOSubsystems.VAD,
+        })
         isSpeech.value = true
         options?.onSpeechStart?.()
       })
 
+      vad.value.on('speech-audio', (event) => {
+        options?.onSpeechAudio?.(event)
+      })
+
       vad.value.on('speech-end', () => {
         isSpeech.value = false
         options?.onSpeechEnd?.()
       })
 
       vad.value.on('speech-ready', (event) => {
+        activeSpan?.setAttribute(IOAttributes.VADAudioDurationMs, event.duration)
+        finishActiveSpan(false)
         options?.onSpeechReady?.(event)
       })
 
@@ -150,6 +174,7 @@ export function useVAD(workerUrl: string, options?: UseVADOptions) {
   }
 
   function dispose() {
+    finishActiveSpan(true)
     manager.value?.stop()
     manager.value?.dispose()
     manager.value = undefined
diff --git a/packages/stage-ui/src/stores/modules/hearing.ts b/packages/stage-ui/src/stores/modules/hearing.ts
index 4388bd329..1c2c28bcc 100644
--- a/packages/stage-ui/src/stores/modules/hearing.ts
+++ b/packages/stage-ui/src/stores/modules/hearing.ts
@@ -15,11 +15,13 @@ import vadWorkletUrl from '../../workers/vad/process.worklet?worker&url'
 
 import { useAnalytics } from '../../composables/use-analytics'
 import { activeTurnSpan, startSpan } from '../../composables/use-io-tracer'
+import { createVadStreamingSession } from '../../libs/audio/vad-streaming-session'
 import { OFFICIAL_TRANSCRIPTION_PROVIDER_ID } from '../../libs/providers'
-import { streamAliyunTranscription } from '../providers/aliyun/stream-transcription'
+import { streamWebSpeechAPITranscription } from '../../libs/providers/providers/browser-web-speech-api'
+import { streamTranscription } from '../../libs/providers/stream-transcription'
+import { useVAD } from '../ai/models/vad'
 import { useProviderConfigStore } from '../providers/config'
 import { useProviderStore } from '../providers/provider'
-import { streamWebSpeechAPITranscription } from '../providers/web-speech-api'
 
 function errorMessage(err: unknown): string {
   const msg = errorMessageFromValue(err)
@@ -39,7 +41,7 @@ function errorMessage(err: unknown): string {
 // an inactive stream. Those cases should not be surfaced as provider failures because the session was
 // explicitly asked to stop. If a future abort is noisy or unexpected, inspect the abort source first:
 // `stopStreamingTranscription()` in this file is the primary origin, and provider-specific teardown
-// bridges such as `packages/stage-ui/src/stores/providers/aliyun/stream-transcription.ts` propagate the
+// provider adapters in `packages/stage-ui/src/libs/providers/providers/` propagate the
 // same reason through the transport. Only treat an abort as "expected" if it is one of these known
 // shutdown paths; any other `AbortError` should still be investigated as a real lifecycle bug or a
 // provider/runtime failure.
@@ -110,6 +112,14 @@ interface HearingTranscriptionInvokeOptions {
   providerOptions?: Record<string, unknown>
 }
 
+interface MediaStreamTranscriptionOptions {
+  sampleRate?: number
+  providerOptions?: Record<string, unknown>
+  idleTimeoutMs?: number
+  onSentenceEnd?: (delta: string) => void
+  onSpeechEnd?: (text: string) => void
+}
+
 export const CONFIDENCE_THRESHOLD_DISABLED = -3
 
 export function filterTranscriptionByConfidence(
@@ -229,8 +239,8 @@ export function resolveTranscriptionFileName(file: File, explicitFileName?: stri
 }
 
 const STREAM_TRANSCRIPTION_EXECUTORS: Record<string, StreamTranscription> = {
-  'aliyun-nls-transcription': streamAliyunTranscription,
-  [OFFICIAL_TRANSCRIPTION_PROVIDER_ID]: streamAliyunTranscription,
+  'aliyun-nls-transcription': streamTranscription,
+  [OFFICIAL_TRANSCRIPTION_PROVIDER_ID]: streamTranscription,
   // Web Speech API is handled specially in transcribeForMediaStream since it works directly with MediaStream
 }
 
@@ -572,9 +582,9 @@ export const useHearingSpeechInputPipeline = defineStore('modules:hearing:speech
     trackVoiceInputStarted,
   } = useAnalytics()
   const streamingSession = shallowRef<{
-    audioContext: AudioContext | Record<string, never>
-    workletNode: AudioWorkletNode | Record<string, never>
-    mediaStreamSource: MediaStreamAudioSourceNode | Record<string, never>
+    audioContext?: AudioContext
+    workletNode?: AudioWorkletNode
+    mediaStreamSource?: MediaStreamAudioSourceNode
     audioStreamController?: ReadableStreamDefaultController<ArrayBuffer>
     abortController: AbortController
     result?: HearingTranscriptionResult & { recognition?: any }
@@ -585,6 +595,19 @@ export const useHearingSpeechInputPipeline = defineStore('modules:hearing:speech
       onSpeechEnd?: (text: string) => void
     }
   }>()
+  const streamingVadSession = shallowRef<{
+    vad: Pick<ReturnType<typeof useVAD>, 'dispose'>
+    lifecycle: ReturnType<typeof createVadStreamingSession>
+    providerId: string
+    callbacks: {
+      onSentenceEnd?: (delta: string) => void
+      onSpeechEnd?: (text: string) => void
+    }
+    activeSegment?: {
+      audioChunks: ArrayBuffer[]
+      audioStreamController?: ReadableStreamDefaultController<ArrayBuffer>
+    }
+  }>()
 
   let asrSpan: Span | undefined
 
@@ -620,66 +643,9 @@ export const useHearingSpeechInputPipeline = defineStore('modules:hearing:speech
     return providersStore.getTranscriptionFeatures(providerId).supportsStreamInput
   })
 
-  const DEFAULT_SAMPLE_RATE = 16000
   const DEFAULT_STREAM_IDLE_TIMEOUT = 15000
 
-  function float32ToInt16(buffer: Float32Array) {
-    const output = new Int16Array(buffer.length)
-    for (let i = 0; i < buffer.length; i++) {
-      const value = Math.max(-1, Math.min(1, buffer[i]))
-      output[i] = value < 0 ? value * 0x8000 : value * 0x7FFF
-    }
-
-    return output
-  }
-
-  async function createAudioStreamFromMediaStream(stream: MediaStream, sampleRate = DEFAULT_SAMPLE_RATE, onActivity?: () => void) {
-    const audioContext = new AudioContext({ sampleRate, latencyHint: 'interactive' })
-    await audioContext.audioWorklet.addModule(vadWorkletUrl)
-    const workletNode = new AudioWorkletNode(audioContext, 'vad-audio-worklet-processor')
-
-    let audioStreamController: ReadableStreamDefaultController<ArrayBuffer> | undefined
-    const audioStream = new ReadableStream<ArrayBuffer>({
-      start(controller) {
-        audioStreamController = controller
-      },
-      cancel: () => {
-        audioStreamController = undefined
-      },
-    })
-
-    workletNode.port.onmessage = ({ data }: MessageEvent<{ buffer?: Float32Array }>) => {
-      const buffer = data?.buffer
-      if (!buffer || !audioStreamController)
-        return
-
-      const pcm16 = float32ToInt16(buffer)
-      // Clone buffer to avoid retaining underlying ArrayBuffer references
-      audioStreamController.enqueue(pcm16.buffer.slice(0))
-      onActivity?.()
-    }
-
-    const mediaStreamSource = audioContext.createMediaStreamSource(stream)
-    mediaStreamSource.connect(workletNode)
-
-    // Sink to avoid feedback/echo
-    const silentGain = audioContext.createGain()
-    silentGain.gain.value = 0
-    workletNode.connect(silentGain)
-    silentGain.connect(audioContext.destination)
-
-    return {
-      audioContext,
-      workletNode,
-      mediaStreamSource,
-      audioStream,
-      get controller() {
-        return audioStreamController
-      },
-    }
-  }
-
-  async function stopStreamingTranscription(abort?: boolean, disposeProviderId?: string) {
+  async function stopRealtimeTranscription(abort?: boolean, disposeProviderId?: string) {
     const session = streamingSession.value
     if (!session)
       return
@@ -749,12 +715,14 @@ export const useHearingSpeechInputPipeline = defineStore('modules:hearing:speech
     }
     catch {}
 
-    await tryCatch(() => {
-      session.mediaStreamSource.disconnect()
-      session.workletNode.port.onmessage = null
-      session.workletNode.disconnect()
-    })
-    await tryCatch(() => session.audioContext.close())
+    if (session.mediaStreamSource && session.workletNode && session.audioContext) {
+      await tryCatch(() => {
+        session.mediaStreamSource?.disconnect()
+        session.workletNode!.port.onmessage = null
+        session.workletNode?.disconnect()
+      })
+      await tryCatch(() => session.audioContext?.close())
+    }
 
     if (session.idleTimer)
       clearTimeout(session.idleTimer)
@@ -787,13 +755,225 @@ export const useHearingSpeechInputPipeline = defineStore('modules:hearing:speech
     return text
   }
 
-  async function transcribeForMediaStream(stream: MediaStream, options?: {
-    sampleRate?: number
-    providerOptions?: Record<string, unknown>
-    idleTimeoutMs?: number
-    onSentenceEnd?: (delta: string) => void
-    onSpeechEnd?: (text: string) => void
-  }) {
+  /** Finishes one VAD segment without aborting the Provider's final response. */
+  async function finishRealtimeTranscription() {
+    const session = streamingSession.value
+    if (!session)
+      return
+
+    try {
+      session.audioStreamController?.close()
+    }
+    catch {}
+
+    if (session.result?.mode !== 'stream') {
+      streamingSession.value = undefined
+      return session.result?.text
+    }
+
+    try {
+      return await session.result.text
+    }
+    catch (err) {
+      if (!isExpectedStreamStopError(err)) {
+        error.value = errorMessage(err)
+        console.error('Error finishing transcription:', error.value)
+      }
+    }
+    finally {
+      if (streamingSession.value === session)
+        streamingSession.value = undefined
+    }
+  }
+
+  /** Stops the active VAD detector and any realtime transcription session. */
+  async function stopStreamingTranscription(abort?: boolean, disposeProviderId?: string) {
+    const vadSession = streamingVadSession.value
+    if (vadSession) {
+      streamingVadSession.value = undefined
+      vadSession.vad.dispose()
+      await vadSession.lifecycle.dispose()
+    }
+
+    return await stopRealtimeTranscription(abort, disposeProviderId)
+  }
+
+  function float32ToInt16(buffer: Float32Array) {
+    const output = new Int16Array(buffer.length)
+    for (let i = 0; i < buffer.length; i++) {
+      const value = Math.max(-1, Math.min(1, buffer[i]))
+      output[i] = value < 0 ? value * 0x8000 : value * 0x7FFF
+    }
+
+    return output
+  }
+
+  function enqueueVadAudio(segment: NonNullable<typeof streamingVadSession.value>['activeSegment'], buffer: Float32Array) {
+    if (!segment)
+      return
+
+    const pcm16 = float32ToInt16(buffer)
+    const chunk = pcm16.buffer.slice(0)
+    if (segment.audioStreamController) {
+      segment.audioStreamController.enqueue(chunk)
+      return
+    }
+
+    segment.audioChunks.push(chunk)
+  }
+
+  function createVadAudioStream(segment: NonNullable<typeof streamingVadSession.value>['activeSegment']) {
+    if (!segment)
+      throw new Error('VAD did not create an active speech segment.')
+
+    return new ReadableStream<ArrayBuffer>({
+      start(controller) {
+        segment.audioStreamController = controller
+        for (const chunk of segment.audioChunks)
+          controller.enqueue(chunk)
+        segment.audioChunks.length = 0
+      },
+      cancel() {
+        segment.audioStreamController = undefined
+        segment.audioChunks.length = 0
+      },
+    })
+  }
+
+  function consumeRealtimeTranscriptionResult(
+    session: NonNullable<typeof streamingSession.value>,
+    result: HearingTranscriptionResult,
+  ) {
+    if (result.mode !== 'stream' || !result.textStream)
+      return
+
+    const sessionSpan = asrSpan
+    const sessionCallbacks = session.callbacks
+    void (async () => {
+      let fullText = ''
+      try {
+        const reader = result.textStream.getReader()
+
+        while (true) {
+          const { done, value } = await reader.read()
+          if (done)
+            break
+          if (!value)
+            continue
+
+          fullText += value
+          sessionSpan?.addEvent(IOEvents.ASRSentenceEnd, { [IOAttributes.ASRText]: value })
+          sessionCallbacks?.onSentenceEnd?.(value)
+        }
+      }
+      catch (err) {
+        if (!isExpectedStreamStopError(err))
+          console.error('Error reading text stream:', err)
+      }
+      finally {
+        sessionSpan?.setAttribute(IOAttributes.ASRText, fullText)
+        sessionSpan?.end()
+        if (asrSpan === sessionSpan)
+          asrSpan = undefined
+        sessionCallbacks?.onSpeechEnd?.(fullText)
+      }
+    })()
+  }
+
+  async function startVadRealtimeTranscription(
+    providerId: string,
+    options: MediaStreamTranscriptionOptions | undefined,
+    vadSession: NonNullable<typeof streamingVadSession.value>,
+  ) {
+    const segment = vadSession.activeSegment
+    if (!segment)
+      return
+
+    const provider = await providersStore.getProviderInstance<TranscriptionProviderWithExtraOptions<string, any>>(providerId)
+    if (!provider)
+      throw new Error('Failed to initialize speech provider')
+
+    const abortController = new AbortController()
+    const session: NonNullable<typeof streamingSession.value> = {
+      audioStreamController: undefined as ReadableStreamDefaultController<ArrayBuffer> | undefined,
+      abortController,
+      providerId,
+      callbacks: vadSession.callbacks,
+    }
+    const audioStream = createVadAudioStream(segment)
+    session.audioStreamController = segment.audioStreamController
+    streamingSession.value = session
+    startStreamingAsrSpan(providerId)
+
+    const result = await hearingStore.transcription(
+      providerId,
+      provider,
+      activeTranscriptionModel.value,
+      { inputAudioStream: audioStream },
+      undefined,
+      {
+        providerOptions: {
+          abortSignal: abortController.signal,
+          ...options?.providerOptions,
+        },
+      },
+    )
+
+    if (streamingSession.value !== session)
+      return
+
+    session.result = result
+    consumeRealtimeTranscriptionResult(session, result)
+  }
+
+  async function startVadStreamingTranscription(
+    stream: MediaStream,
+    providerId: string,
+    options: MediaStreamTranscriptionOptions | undefined,
+  ) {
+    let vadSession!: NonNullable<typeof streamingVadSession.value>
+    const vad = useVAD(vadWorkletUrl, {
+      onSpeechStart: () => {
+        vadSession.activeSegment = { audioChunks: [] }
+        vadSession.lifecycle.onSpeechStart()
+      },
+      onSpeechAudio: ({ buffer }) => {
+        enqueueVadAudio(vadSession.activeSegment, buffer)
+      },
+      onSpeechEnd: () => {
+        vadSession.lifecycle.onSpeechEnd()
+      },
+    })
+    const lifecycle = createVadStreamingSession({
+      start: async () => await startVadRealtimeTranscription(providerId, options, vadSession),
+      stop: async () => {
+        await finishRealtimeTranscription()
+      },
+      onError: (err) => {
+        error.value = errorMessage(err)
+        console.error('Error managing VAD streaming transcription:', error.value)
+      },
+    })
+    vadSession = {
+      vad,
+      lifecycle,
+      providerId,
+      callbacks: {
+        onSentenceEnd: options?.onSentenceEnd,
+        onSpeechEnd: options?.onSpeechEnd,
+      },
+    }
+    streamingVadSession.value = vadSession
+
+    await vad.init()
+    if (!vad.loaded.value) {
+      throw new Error(vad.inferenceError.value || 'Failed to initialize voice activity detection.')
+    }
+
+    await vad.start(stream)
+  }
+
+  async function transcribeForMediaStream(stream: MediaStream, options?: MediaStreamTranscriptionOptions) {
     console.info('[Hearing Pipeline] transcribeForMediaStream called', {
       supportsStreamInput: supportsStreamInput.value,
       hasStream: !!stream,
@@ -974,136 +1154,24 @@ export const useHearingSpeechInputPipeline = defineStore('modules:hearing:speech
         return
       }
 
-      const provider = await providersStore.getProviderInstance<TranscriptionProviderWithExtraOptions<string, any>>(providerId)
-      if (!provider) {
-        throw new Error('Failed to initialize speech provider')
-      }
-
-      const idleTimeout = options?.idleTimeoutMs ?? DEFAULT_STREAM_IDLE_TIMEOUT
-
-      // If a session exists, reuse it unless new callbacks are provided.
-      // The stream reader captures callbacks at creation time, so updated callbacks
-      // require restarting the session to create a new reader.
-      const existingSession = streamingSession.value
-      if (existingSession) {
-        const hasNewCallbacks = haveStreamingCallbacksChanged(existingSession.callbacks, {
+      const existingVadSession = streamingVadSession.value
+      if (existingVadSession) {
+        const hasNewCallbacks = haveStreamingCallbacksChanged(existingVadSession.callbacks, {
           onSentenceEnd: options?.onSentenceEnd,
           onSpeechEnd: options?.onSpeechEnd,
         })
 
         if (hasNewCallbacks) {
-          console.info('[Hearing Pipeline] New callbacks provided, restarting session')
-          await stopStreamingTranscription(false, existingSession.providerId)
-          // Fall through to create a new session with updated callbacks
+          console.info('[Hearing Pipeline] New callbacks provided, restarting VAD detection')
+          await stopStreamingTranscription(false, existingVadSession.providerId)
         }
         else {
-          // No callback changes: refresh idle timer and reuse session
-          if (existingSession.idleTimer) {
-            clearTimeout(existingSession.idleTimer)
-            existingSession.idleTimer = setTimeout(async () => {
-              await stopStreamingTranscription(false, existingSession.providerId)
-            }, idleTimeout)
-          }
+          console.info('[Hearing Pipeline] VAD detection already active, reusing it')
           return
         }
       }
 
-      startStreamingAsrSpan(providerId)
-
-      const abortController = new AbortController()
-      let idleTimer: ReturnType<typeof setTimeout> | undefined
-      const bumpIdle = () => {
-        if (idleTimer)
-          clearTimeout(idleTimer)
-        idleTimer = setTimeout(async () => {
-          await stopStreamingTranscription(false, providerId)
-        }, idleTimeout)
-      }
-
-      const session = await createAudioStreamFromMediaStream(
-        stream,
-        options?.sampleRate ?? DEFAULT_SAMPLE_RATE,
-        () => bumpIdle(),
-      )
-
-      if (session.audioContext.state === 'suspended')
-        await session.audioContext.resume()
-
-      bumpIdle()
-
-      const model = activeTranscriptionModel.value
-      const result = await hearingStore.transcription(
-        providerId,
-        provider,
-        model,
-        { inputAudioStream: session.audioStream },
-        undefined,
-        {
-          providerOptions: {
-            abortSignal: abortController.signal,
-            ...options?.providerOptions,
-          },
-        },
-      )
-
-      streamingSession.value = {
-        audioContext: session.audioContext,
-        workletNode: session.workletNode,
-        mediaStreamSource: session.mediaStreamSource,
-        audioStreamController: session.controller,
-        abortController,
-        result,
-        idleTimer,
-        providerId,
-        callbacks: {
-          onSentenceEnd: options?.onSentenceEnd,
-          onSpeechEnd: options?.onSpeechEnd,
-        },
-      }
-
-      // Stream out text deltas to caller without tearing down the session.
-      if (result.mode === 'stream' && result.textStream) {
-        void (async () => {
-          // Capture callbacks from the session at the time the reader is created
-          // This prevents cross-session leakage if the session is restarted before
-          // this reader finishes (e.g., when navigating between pages or callbacks change)
-          const sessionCallbacks = {
-            onSentenceEnd: streamingSession.value?.callbacks?.onSentenceEnd,
-            onSpeechEnd: streamingSession.value?.callbacks?.onSpeechEnd,
-          }
-
-          let fullText = ''
-          try {
-            const reader = result.textStream.getReader()
-
-            while (true) {
-              const { done, value } = await reader.read()
-              if (done)
-                break
-              if (value) {
-                fullText += value
-                if (asrSpan)
-                  asrSpan.addEvent(IOEvents.ASRSentenceEnd, { [IOAttributes.ASRText]: value })
-                // Use captured callbacks to avoid cross-session leakage
-                sessionCallbacks.onSentenceEnd?.(value)
-              }
-            }
-          }
-          catch (err) {
-            if (!isExpectedStreamStopError(err))
-              console.error('Error reading text stream:', err)
-          }
-          finally {
-            if (asrSpan) {
-              asrSpan.setAttribute(IOAttributes.ASRText, fullText)
-              asrSpan.end()
-              asrSpan = undefined
-            }
-            // Use captured callbacks to avoid cross-session leakage
-            sessionCallbacks.onSpeechEnd?.(fullText)
-          }
-        })()
-      }
+      await startVadStreamingTranscription(stream, providerId, options)
     }
     catch (err) {
       endStreamingAsrSpan()
diff --git a/packages/stage-ui/src/workers/vad/vad.ts b/packages/stage-ui/src/workers/vad/vad.ts
index a7e63ab76..eeb1aa387 100644
--- a/packages/stage-ui/src/workers/vad/vad.ts
+++ b/packages/stage-ui/src/workers/vad/vad.ts
@@ -147,6 +147,10 @@ export class VAD implements BaseVAD {
         // Speech just started
         this.emit('speech-start', undefined)
         this.emit('status', { type: 'info', message: 'Speech detected' })
+        this.emit('speech-audio', { buffer: this.createLeadingSpeechAudio(inputBuffer) })
+      }
+      else {
+        this.emit('speech-audio', { buffer: inputBuffer.slice() })
       }
 
       // Update state
@@ -157,6 +161,7 @@ export class VAD implements BaseVAD {
     }
 
     // At this point, we were recording but the current buffer is not speech
+    this.emit('speech-audio', { buffer: inputBuffer.slice() })
     this.postSpeechSamples += inputBuffer.length
 
     // Check if silence is long enough to consider speech ended
@@ -164,6 +169,7 @@ export class VAD implements BaseVAD {
       // Check if the speech segment is long enough to process
       if (this.bufferPointer < minSpeechDurationSamples) {
         // Too short, reset without processing
+        this.emit('speech-end', undefined)
         this.reset()
 
         return
@@ -174,6 +180,21 @@ export class VAD implements BaseVAD {
     }
   }
 
+  /** Combines VAD's retained pre-speech padding with the first detected speech chunk. */
+  private createLeadingSpeechAudio(inputBuffer: Float32Array): Float32Array {
+    const leadingLength = this.prevBuffers.reduce((total, buffer) => total + buffer.length, 0)
+    const output = new Float32Array(leadingLength + inputBuffer.length)
+    let offset = 0
+
+    for (const buffer of this.prevBuffers) {
+      output.set(buffer, offset)
+      offset += buffer.length
+    }
+
+    output.set(inputBuffer, offset)
+    return output
+  }
+
   /**
    * Detect speech in an audio buffer
    */
diff --git a/server/apps/api/src/routes/audio-transcription-stream/session.ts b/server/apps/api/src/routes/audio-transcription-stream/session.ts
index ec50070d8..e4d07f9d1 100644
--- a/server/apps/api/src/routes/audio-transcription-stream/session.ts
+++ b/server/apps/api/src/routes/audio-transcription-stream/session.ts
@@ -169,7 +169,7 @@ async function writeAudioToUpstream(audioStream: ReadableStream<Uint8Array>, ws:
  * - `audioStream` contains 16 kHz PCM chunks by default, matching the Hearing worklet output.
  *
  * Returns:
- * - A `text/event-stream` response consumable by the existing `streamAliyunTranscription` executor.
+ * - A `text/event-stream` response consumable by the shared `streamTranscription` adapter.
  */
 export function createAliyunNlsStreamResponse(options: CreateAliyunNlsStreamResponseOptions): Response {
   const body = new ReadableStream<Uint8Array>({