fix: 修复语音消息取错、时长精度与转写不刷新

- 时长解析：改取 <voicemsg voicelength>（毫秒），不再误取 length（SILK 字节数） - 时长显示：四舍五入对齐微信口径，有原生时长时不被解码时长覆盖 - 转写：重新识别改为强制重算，跳过身份级缓存短路（音频级缓存仍生效） - 日报：语音累计秒数显示取整，不再出现小数
2026-10-03 18:33:14 +08:00 · 2026-09-20 15:40:01 +08:00
parent 9b82d29037
commit 9593ca0f54
21 changed files with 435 additions and 52 deletions
@@ -1,4 +1,4 @@
-import { render, screen, waitFor } from '@testing-library/react'
+import { act, render, screen, waitFor } from '@testing-library/react'
 import userEvent from '@testing-library/user-event'
 import { beforeEach, describe, expect, it, vi } from 'vitest'
 import { VoicePlayer } from '../../src/renderer/src/components/VoicePlayer'
@@ -6,6 +6,8 @@ import { VoicePlayer } from '../../src/renderer/src/components/VoicePlayer'
 const play = vi.fn(() => Promise.resolve())
 const pause = vi.fn()

+let lastAudio: FakeAudio | null = null
+
 class FakeAudio {
  preload = ''
  src = ''
@@ -18,12 +20,27 @@ class FakeAudio {
  pause = pause
  load = vi.fn()
  removeAttribute = vi.fn()
+
+  constructor() {
+    lastAudio = this
+  }
 }

+const renderPlayer = (duration?: number) =>
+  render(
+    <VoicePlayer
+      sessionId="filehelper"
+      localId={11}
+      createTime={1785553200}
+      duration={duration}
+    />
+  )
+
 describe('VoicePlayer', () => {
  beforeEach(() => {
    play.mockClear()
    pause.mockClear()
+    lastAudio = null
    vi.stubGlobal('Audio', FakeAudio)
    window.api = {
      getVoiceData: vi.fn().mockResolvedValue({
@@ -70,17 +87,73 @@ describe('VoicePlayer', () => {
    render(<VoicePlayer sessionId="filehelper" localId={11} createTime={1785553200} duration={1} />)
    await userEvent.click(screen.getByRole('button', { name: '转文字' }))

+    // 用户主动触发必须带 force：否则会被身份级缓存（不含 audio_hash）短路。
    await waitFor(() =>
-      expect(window.api.recognizeVoice).toHaveBeenCalledWith({
-        sessionId: 'filehelper',
-        localId: 11,
-        createTime: 1785553200,
-        svrId: undefined
-      })
+      expect(window.api.recognizeVoice).toHaveBeenCalledWith(
+        {
+          sessionId: 'filehelper',
+          localId: 11,
+          createTime: 1785553200,
+          svrId: undefined
+        },
+        { force: true }
+      )
    )
    expect(await screen.findByText('这是固定的测试转写')).toBeInTheDocument()
  })

+  it('rounds the shown duration the way WeChat does', () => {
+    // 微信 4211ms 显示 4"，所以口径是 round 不是 floor；进位要传给分钟位。
+    const cases: Array<[number | undefined, string]> = [
+      [1.7, '0:02'],
+      [4.211, '0:04'],
+      [7.505, '0:08'],
+      [59.6, '1:00'],
+      [119.6, '2:00'],
+      [3, '0:03'],
+      [7, '0:07'],
+      [15, '0:15'],
+      [0, '0:00'],
+      [undefined, '0:00'],
+      [NaN, '0:00']
+    ]
+    for (const [duration, expected] of cases) {
+      const { container, unmount } = renderPlayer(duration)
+      expect(container.querySelector('.voice-duration')?.textContent).toBe(expected)
+      unmount()
+    }
+  })
+
+  it('keeps the WeChat duration instead of the systematically short decoded one', async () => {
+    const { container } = renderPlayer(1.979)
+    expect(container.querySelector('.voice-duration')?.textContent).toBe('0:02')
+
+    await userEvent.click(container.querySelector('.voice-message') as HTMLElement)
+    await waitFor(() => expect(lastAudio).not.toBeNull())
+    // Silk 解码时长比微信 length 系统性偏短（实测少 20–279ms），用它覆盖会把精度弄丢。
+    lastAudio!.duration = 1
+    await act(async () => {
+      lastAudio!.onloadedmetadata?.()
+      lastAudio!.ontimeupdate?.()
+    })
+
+    expect(container.querySelector('.voice-duration')?.textContent).toBe('0:02')
+  })
+
+  it('falls back to the decoded duration when WeChat did not provide one', async () => {
+    const { container } = renderPlayer()
+    expect(container.querySelector('.voice-duration')?.textContent).toBe('0:00')
+
+    await userEvent.click(container.querySelector('.voice-message') as HTMLElement)
+    await waitFor(() => expect(lastAudio).not.toBeNull())
+    lastAudio!.duration = 59.6
+    lastAudio!.onloadedmetadata?.()
+
+    await waitFor(() =>
+      expect(container.querySelector('.voice-duration')?.textContent).toBe('1:00')
+    )
+  })
+
  it('opens centralized settings when recognition assets are missing', async () => {
    vi.mocked(window.api.getVoiceModelStatus).mockResolvedValue({
      modelId: 'sensevoice-small-int8',
@@ -4,7 +4,10 @@ import {
  getSummaryDateRangeAt,
  parseGroupDailyReport
 } from '../../src/renderer/src/utils/group-report'
-import { summaryContent } from '../../src/renderer/src/utils/group-report-facts'
+import {
+  buildGroupReportFacts,
+  summaryContent
+} from '../../src/renderer/src/utils/group-report-facts'
 import { summarySender } from '../../src/renderer/src/utils/group-report-facts'
 import { selectHeroParticipantNames } from '../../src/shared/group-report'
 import type { GroupReportMetadata } from '../../src/shared/group-report'
@@ -62,6 +65,37 @@ describe('group report parsing', () => {
    expect(summaryContent(message)).toContain('今晚八点确认发布。')
  })

+  it('rounds accumulated voice seconds in the facts summary', async () => {
+    // 语音时长是 <voicemsg voicelength>（毫秒）换算来的小数秒（1600ms → 1.6）。
+    // 累加必须保留原始精度（否则多条累积会越差越多），但展示给用户的文案要取整。
+    const messages: Message[] = [
+      {
+        id: 'voice-a',
+        from: 'member',
+        type: '语音',
+        datetime: '2026-08-06 10:00:00',
+        content: '[语音]',
+        isSender: false,
+        contentData: { type: 'voice', duration: 1.979 }
+      },
+      {
+        id: 'voice-b',
+        from: 'member',
+        type: '语音',
+        datetime: '2026-08-06 10:00:05',
+        content: '[语音]',
+        isSender: false,
+        contentData: { type: 'voice', duration: 5.379 }
+      }
+    ]
+
+    const snapshot = await buildGroupReportFacts(messages, null, true, 'compact')
+
+    // 1.979 + 5.379 = 7.358 → 文案必须显示整数
+    expect(snapshot.factsPrompt).toContain('累计 7 秒')
+    expect(snapshot.factsPrompt).not.toContain('7.358')
+  })
+
  it('includes a voice transcript when legacy cached messages have no contentData', () => {
    const message: Message = {
      id: 'voice-legacy',
@@ -17,6 +17,38 @@ describe('message parser', () => {
    ).toMatchObject({ type: 'sticker', md5: 'abcdefabcdefabcdefabcdefabcdefab' })
  })

+  it('reads the WeChat voice length in milliseconds and keeps fractional seconds', () => {
+    // 真机实测（2026-09-20）：voicelength 才是毫秒时长，length 是编码数据长度——别取错。
+    // 属性值取自一条真机采样的语音（1.6 秒，微信气泡显示 2"）。
+    const parsed = parseMessageContent(
+      '<msg><voicemsg endflag="1" cancelflag="0" forwardflag="0" voiceformat="4" voicelength="1600" length="6672" bufid="0" /></msg>',
+      34
+    )
+    expect(parsed).toEqual({ type: 'voice', duration: 1.6 })
+    // 误取 length 会得到 6.672 秒（把 2" 的语音显示成 0:07）——这条断言就是防这个回归。
+    expect(parsed).not.toEqual({ type: 'voice', duration: 6.672 })
+
+    // 微信四舍五入到整秒，取整必须在显示层做，不能在解析层丢精度。
+    expect(parseMessageContent('<msg><voicemsg voicelength="4211" /></msg>', 34)).toEqual({
+      type: 'voice',
+      duration: 4.211
+    })
+  })
+
+  it('leaves the voice duration undefined when the payload is missing or unusable', () => {
+    expect(parseMessageContent('', 34)).toEqual({ type: 'voice' })
+    expect(parseMessageContent('voice fixture', 34)).toEqual({ type: 'voice' })
+    expect(parseMessageContent('<msg><voicemsg voiceformat="4" /></msg>', 34)).toEqual({
+      type: 'voice'
+    })
+    expect(parseMessageContent('<msg><voicemsg voicelength="0" /></msg>', 34)).toEqual({
+      type: 'voice'
+    })
+    expect(parseMessageContent('<msg><voicemsg voicelength="abc" /></msg>', 34)).toEqual({
+      type: 'voice'
+    })
+  })
+
  it('keeps video metadata when WeChat omits every MD5 field', () => {
    const parsed = parseMessageContent(
      '<msg><videomsg length="6402169" playlength="30" cdnthumbwidth="224" cdnthumbheight="398" aeskey="25201cc658042689d1ad6747cea2b240" rawmd5="" /></msg>',
@@ -361,3 +361,128 @@ describe('voice pipeline cache lookup', () => {
    repository.close()
  })
 })
+
+/*
+ * force 的语义：用户主动触发时必须绕开 findCompatible（它只按消息身份匹配，
+ * 不含 audio_hash），重新取一次音频；音频级 find(key) 带 audio_hash，保留。
+ */
+describe('voice pipeline force recognition', () => {
+  const forceRoot = mkdtempSync(join(tmpdir(), 'wxe-voice-pipeline-force-'))
+  afterAll(() => rmSync(forceRoot, { recursive: true, force: true }))
+
+  function createPipeline(repository: SqliteTranscriptRepository, transcript: string) {
+    const resolve = vi.fn().mockResolvedValue({
+      data: Buffer.from('encoded'),
+      codec: 'silk',
+      sourceHash: 'fresh-audio'
+    })
+    const decode = vi.fn().mockResolvedValue({
+      pcm: Buffer.from([1, 0]),
+      sampleRate: 16000,
+      channels: 1,
+      sourceHash: 'fresh-audio'
+    })
+    const process = vi.fn().mockReturnValue({
+      samples: new Float32Array([0.1]),
+      sampleRate: 16000,
+      channels: 1,
+      sourceHash: 'fresh-audio',
+      processorVersion: 'processor-v1',
+      durationMs: 1
+    })
+    const recognize = vi.fn().mockResolvedValue({ text: transcript })
+    const pipeline = new VoicePipeline(
+      { resolve },
+      { decode } as never,
+      { version: 'processor-v1', process },
+      {
+        metadata: {
+          recognizerId: 'sensevoice',
+          modelVersion: 'model-v1',
+          modelFingerprint: 'fingerprint-a'
+        },
+        recognize,
+        dispose: vi.fn()
+      },
+      repository
+    )
+    return { pipeline, resolve, decode, process, recognize }
+  }
+
+  function seedStaleTranscript(repository: SqliteTranscriptRepository) {
+    repository.save({
+      accountId: 'account-a',
+      messageIdentity: voiceMessageIdentity(reference),
+      audioHash: 'stale-audio',
+      processorVersion: 'processor-v1',
+      recognizerId: 'sensevoice',
+      modelVersion: 'model-v1',
+      modelFingerprint: 'fingerprint-a',
+      transcript: '旧音频的转写',
+      durationMs: 900,
+      createdAt: 1,
+      updatedAt: 1
+    })
+  }
+
+  const reference = { sessionId: 'session', localId: 1, createTime: 2 }
+
+  it('skips the identity-level cache when force is requested', async () => {
+    const repository = new SqliteTranscriptRepository(join(forceRoot, 'forced.sqlite'))
+    seedStaleTranscript(repository)
+    const findCompatible = vi.spyOn(repository, 'findCompatible')
+    const find = vi.spyOn(repository, 'find')
+    const { pipeline, resolve, recognize } = createPipeline(repository, '正确音频的转写')
+
+    await expect(
+      pipeline.run('account-a', reference, undefined, { force: true })
+    ).resolves.toMatchObject({ transcript: '正确音频的转写', cached: false })
+
+    expect(findCompatible).not.toHaveBeenCalled()
+    expect(resolve).toHaveBeenCalledOnce()
+    expect(find).toHaveBeenCalled()
+    expect(recognize).toHaveBeenCalledOnce()
+    expect(
+      repository.find({
+        accountId: 'account-a',
+        messageIdentity: voiceMessageIdentity(reference),
+        audioHash: 'fresh-audio',
+        processorVersion: 'processor-v1',
+        recognizerId: 'sensevoice',
+        modelVersion: 'model-v1',
+        modelFingerprint: 'fingerprint-a'
+      })?.transcript
+    ).toBe('正确音频的转写')
+    repository.close()
+  })
+
+  it('keeps consulting the identity-level cache without force', async () => {
+    const repository = new SqliteTranscriptRepository(join(forceRoot, 'unchanged.sqlite'))
+    seedStaleTranscript(repository)
+    const findCompatible = vi.spyOn(repository, 'findCompatible')
+    const { pipeline, resolve, decode, process, recognize } = createPipeline(
+      repository,
+      '新转写'
+    )
+
+    await expect(pipeline.run('account-a', reference)).resolves.toMatchObject({
+      transcript: '旧音频的转写',
+      cached: true
+    })
+    expect(findCompatible).toHaveBeenCalledOnce()
+    expect(resolve).not.toHaveBeenCalled()
+    expect(decode).not.toHaveBeenCalled()
+    expect(process).not.toHaveBeenCalled()
+    expect(recognize).not.toHaveBeenCalled()
+    repository.close()
+
+    const explicitFalse = new SqliteTranscriptRepository(join(forceRoot, 'explicit-false.sqlite'))
+    seedStaleTranscript(explicitFalse)
+    const second = createPipeline(explicitFalse, '新转写')
+    await expect(
+      second.pipeline.run('account-a', reference, undefined, { force: false })
+    ).resolves.toMatchObject({ transcript: '旧音频的转写', cached: true })
+    expect(second.resolve).not.toHaveBeenCalled()
+    explicitFalse.close()
+  })
+})
@@ -101,6 +101,49 @@ describe('VoiceRecognitionUseCase transcript updates', () => {
    await useCase.dispose()
  })

+  it('forwards force to the pipeline so manual re-recognition can bypass the identity cache', async () => {
+    const useCase = createUseCase()
+    const state = useCase as unknown as {
+      pipeline: { run: ReturnType<typeof vi.fn> }
+    }
+    state.pipeline.run.mockResolvedValue({
+      transcript: '重新识别的文字',
+      durationMs: 900,
+      cached: false
+    })
+    const reference = { sessionId: 'fixture-contact', localId: 12, createTime: 1_785_895_203 }
+
+    const result = await useCase.recognize(reference, { force: true })
+
+    expect(result).toMatchObject({ success: true, transcript: '重新识别的文字' })
+    expect(state.pipeline.run).toHaveBeenCalledWith(
+      'account-a',
+      reference,
+      expect.anything(),
+      { force: true }
+    )
+    await useCase.dispose()
+  })
+
+  it('leaves force undefined when the caller does not ask for a recompute', async () => {
+    const useCase = createUseCase()
+    const state = useCase as unknown as {
+      pipeline: { run: ReturnType<typeof vi.fn> }
+    }
+    state.pipeline.run.mockResolvedValue({ transcript: '缓存文字', durationMs: 900, cached: true })
+    const reference = { sessionId: 'fixture-contact', localId: 13, createTime: 1_785_895_204 }
+
+    await useCase.recognize(reference)
+
+    expect(state.pipeline.run).toHaveBeenCalledWith(
+      'account-a',
+      reference,
+      expect.anything(),
+      { force: undefined }
+    )
+    await useCase.dispose()
+  })
+
  it('publishes an explicit cached transcript for a coalesced export index refresh', async () => {
    const useCase = createUseCase()
    const listener = vi.fn().mockResolvedValue(undefined)