feat(ui,stage-ui,stage-pages,i18n): transcription confidence filter (#1148)

2026-03-30 03:40:49 +08:00
parent 5a11ff7b25
commit 7ad4debe11
7 changed files with 123 additions and 6 deletions
@@ -404,6 +404,14 @@ pages:
        section:
          provider-selection:
            description: Select the suitable speech recognition provider
+          confidence-threshold:
+            title: Confidence Threshold
+            description: Filter out low-confidence transcriptions to reduce Whisper hallucinations. Values closer to 0 are more strict; drag to the leftmost to disable. Only effective for providers supporting Whisper API (e.g., OpenAI, Groq).
+            disabled: Disabled
+            verbose-json-note: >-
+              Note: If your provider does not support verbose_json responses, this setting will have no effect.
+            verbose-json-unsupported: >-
+              Your provider did not return verbose_json segments. Confidence filtering had no effect on the last transcription.
    memory-long-term:
      description: Long-term memory specific settings and management
      title: Long-Term Memory
@@ -388,6 +388,12 @@ pages:
        section:
          provider-selection:
            description: 选择合适的语音转文本的服务来源
+          confidence-threshold:
+            title: 置信度阈值
+            description: 过滤低置信度的转录结果，避免 Whisper 幻觉噪音。值越接近 0 越严格，拖到最左为禁用。仅对支持 Whisper API 的服务商有效（如 OpenAI、Groq）。
+            disabled: 已禁用
+            verbose-json-note: 注意：如果你的服务商不支持 verbose_json 响应，此设置将不会生效。
+            verbose-json-unsupported: 你的服务商未返回 verbose_json 片段，上次转录的置信度过滤未生效。
    memory-long-term:
      description: 长期记忆
      title: 长期记忆
@@ -379,6 +379,12 @@ pages:
        section:
          provider-selection:
            description: 選擇合適的語音辨識提供者
+          confidence-threshold:
+            title: 置信度閾值
+            description: 過濾低置信度的轉錄結果，避免 Whisper 幻覺噪音。值越接近 0 越嚴格，拖到最左為禁用。僅對支援 Whisper API 的服務商有效（如 OpenAI、Groq）。
+            disabled: 已停用
+            verbose-json-note: 注意：如果你的服務商不支援 verbose_json 響應，此設定將不會生效。
+            verbose-json-unsupported: 你的服務商未返回 verbose_json 片段，上次轉錄的置信度過濾未生效。
    memory-long-term:
      description: 長期記憶
      title: 長期記憶
@@ -5,7 +5,7 @@ import { Alert, ErrorContainer, LevelMeter, RadioCardManySelect, RadioCardSimple
 import { useAnalytics, useAudioAnalyzer, useAudioRecorder } from '@proj-airi/stage-ui/composables'
 import { useVAD } from '@proj-airi/stage-ui/stores/ai/models/vad'
 import { useAudioContext } from '@proj-airi/stage-ui/stores/audio'
-import { useHearingSpeechInputPipeline, useHearingStore } from '@proj-airi/stage-ui/stores/modules/hearing'
+import { CONFIDENCE_THRESHOLD_DISABLED, useHearingSpeechInputPipeline, useHearingStore } from '@proj-airi/stage-ui/stores/modules/hearing'
 import { useProvidersStore } from '@proj-airi/stage-ui/stores/providers'
 import { useSettingsAudioDevice } from '@proj-airi/stage-ui/stores/settings'
 import { Button, FieldCheckbox, FieldCombobox, FieldInput, FieldRange } from '@proj-airi/ui'
@@ -28,6 +28,8 @@ const {
  activeCustomModelName,
  autoSendEnabled,
  autoSendDelay,
+  confidenceThreshold,
+  verboseJsonNotSupported,
 } = storeToRefs(hearingStore)
 const providersStore = useProvidersStore()
 const { configuredTranscriptionProvidersMetadata } = storeToRefs(providersStore)
@@ -656,6 +658,32 @@ onUnmounted(() => {
          </div>
        </div>

+        <!-- Confidence threshold (only for non-streaming providers) -->
+        <div v-if="!supportsStreamInput" class="border-t border-neutral-200 pt-4 dark:border-neutral-700">
+          <div class="mb-4">
+            <h2 class="text-lg text-neutral-500 md:text-2xl dark:text-neutral-500">
+              {{ t('settings.pages.modules.hearing.sections.section.confidence-threshold.title') }}
+            </h2>
+            <div text="neutral-400 dark:neutral-400">
+              {{ t('settings.pages.modules.hearing.sections.section.confidence-threshold.description') }}
+            </div>
+          </div>
+          <FieldRange
+            v-model="confidenceThreshold"
+            :min="CONFIDENCE_THRESHOLD_DISABLED"
+            :max="0"
+            :step="0.1"
+            :format-value="value => value <= CONFIDENCE_THRESHOLD_DISABLED ? t('settings.pages.modules.hearing.sections.section.confidence-threshold.disabled') : value.toFixed(1)"
+          />
+          <div v-if="confidenceThreshold > CONFIDENCE_THRESHOLD_DISABLED" class="mt-2 text-xs text-neutral-400 dark:text-neutral-500">
+            {{ t('settings.pages.modules.hearing.sections.section.confidence-threshold.verbose-json-note') }}
+          </div>
+          <div v-if="verboseJsonNotSupported" class="mt-2 flex items-center gap-1.5 text-xs text-amber-500 dark:text-amber-400">
+            <div i-solar:warning-circle-line-duotone class="shrink-0" />
+            {{ t('settings.pages.modules.hearing.sections.section.confidence-threshold.verbose-json-unsupported') }}
+          </div>
+        </div>
+
        <!-- Auto-send settings -->
        <div class="border-t border-neutral-200 pt-4 dark:border-neutral-700">
          <div class="mb-4">
@@ -0,0 +1,31 @@
+import { describe, expect, it } from 'vitest'
+
+import { filterTranscriptionByConfidence } from './hearing'
+
+describe('filterTranscriptionByConfidence', () => {
+  const segments = [
+    { text: 'Hello ', avg_logprob: -0.3 },
+    { text: 'world ', avg_logprob: -1.2 },
+    { text: 'gibberish', avg_logprob: -2.5 },
+  ]
+
+  it('keeps all segments when threshold is very low', () => {
+    expect(filterTranscriptionByConfidence(segments, -3)).toBe('Hello world gibberish')
+  })
+
+  it('filters out low-confidence segments', () => {
+    expect(filterTranscriptionByConfidence(segments, -1)).toBe('Hello')
+  })
+
+  it('filters out all segments when threshold is 0', () => {
+    expect(filterTranscriptionByConfidence(segments, 0)).toBe('')
+  })
+
+  it('returns empty string for empty segments', () => {
+    expect(filterTranscriptionByConfidence([], -1)).toBe('')
+  })
+
+  it('trims whitespace from result', () => {
+    expect(filterTranscriptionByConfidence([{ text: '  hello  ', avg_logprob: -0.5 }], -1)).toBe('hello')
+  })
+})
@@ -7,7 +7,7 @@ import { useLocalStorageManualReset } from '@proj-airi/stage-shared/composables'
 import { refManualReset } from '@vueuse/core'
 import { generateTranscription } from '@xsai/generate-transcription'
 import { defineStore, storeToRefs } from 'pinia'
-import { computed, ref, shallowRef } from 'vue'
+import { computed, ref, shallowRef, watch } from 'vue'

 import vadWorkletUrl from '../../workers/vad/process.worklet?worker&url'

@@ -76,6 +76,19 @@ interface HearingTranscriptionInvokeOptions {
  providerOptions?: Record<string, unknown>
 }

+export const CONFIDENCE_THRESHOLD_DISABLED = -3
+
+export function filterTranscriptionByConfidence(
+  segments: Array<{ text?: string, avg_logprob?: number }>,
+  threshold: number,
+): string {
+  if (!segments.some(s => s?.avg_logprob != null && s?.text != null)) {
+    return ''
+  }
+
+  return segments.filter(s => (s?.avg_logprob ?? -Infinity) >= threshold).map(s => s?.text ?? '').join('').trim()
+}
+
 const STREAM_TRANSCRIPTION_EXECUTORS: Record<string, StreamTranscription> = {
  'aliyun-nls-transcription': streamAliyunTranscription,
  // Web Speech API is handled specially in transcribeForMediaStream since it works directly with MediaStream
@@ -92,6 +105,12 @@ export const useHearingStore = defineStore('hearing-store', () => {
  const transcriptionModelSearchQuery = refManualReset<string>('')
  const autoSendEnabled = useLocalStorageManualReset<boolean>('settings/hearing/auto-send-enabled', false)
  const autoSendDelay = useLocalStorageManualReset<number>('settings/hearing/auto-send-delay', 2000) // Default 2 seconds
+  const confidenceThreshold = useLocalStorageManualReset<number>('settings/hearing/confidence-threshold', CONFIDENCE_THRESHOLD_DISABLED)
+  const verboseJsonNotSupported = ref(false)
+
+  watch(activeTranscriptionProvider, () => {
+    verboseJsonNotSupported.value = false
+  })

  // Computed properties
  const availableProvidersMetadata = computed(() => allAudioTranscriptionProvidersMetadata.value)
@@ -154,6 +173,7 @@ export const useHearingStore = defineStore('hearing-store', () => {
    transcriptionModelSearchQuery.reset()
    autoSendEnabled.reset()
    autoSendDelay.reset()
+    confidenceThreshold.reset()
  }

  async function transcription(
@@ -217,12 +237,28 @@ export const useHearingStore = defineStore('hearing-store', () => {
      throw new Error('File input is required for transcription.')
    }

+    const useVerboseJson = !format && confidenceThreshold.value > CONFIDENCE_THRESHOLD_DISABLED
    const response = await generateTranscription({
      ...provider.transcription(model, options?.providerOptions),
      file: normalizedInput.file,
-      responseFormat: format,
+      responseFormat: useVerboseJson ? 'verbose_json' : format,
    })

+    if (useVerboseJson) {
+      if (response.segments) {
+        verboseJsonNotSupported.value = false
+        return {
+          mode: 'generate',
+          ...response,
+          text: filterTranscriptionByConfidence(response.segments, confidenceThreshold.value),
+        }
+      }
+      else {
+        verboseJsonNotSupported.value = true
+        console.warn('[Hearing] Confidence filter is enabled but the provider did not return verbose_json segments. Filtering has no effect.')
+      }
+    }
+
    return {
      mode: 'generate',
      ...response,
@@ -237,6 +273,8 @@ export const useHearingStore = defineStore('hearing-store', () => {
    transcriptionModelSearchQuery,
    autoSendEnabled,
    autoSendDelay,
+    confidenceThreshold,
+    verboseJsonNotSupported,

    supportsModelListing,
    providerModels,
@@ -36,9 +36,9 @@ const modelValue = defineModel<number>({ required: true })
    <div :class="['flex', 'flex-row', 'items-center', 'gap-2']">
      <Range
        v-model="modelValue"
-        :min="min || 0"
-        :max="max || 1"
-        :step="step || 0.01"
+        :min="min ?? 0"
+        :max="max ?? 1"
+        :step="step ?? 0.01"
        :class="['w-full']"
      />
    </div>