diff --git a/packages/stage-ui/src/stores/modules/hearing.test.ts b/packages/stage-ui/src/stores/modules/hearing.test.ts
new file mode 100644
index 000000000..654717208
--- /dev/null
+++ b/packages/stage-ui/src/stores/modules/hearing.test.ts
@@ -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')
+ })
+})
diff --git a/packages/stage-ui/src/stores/modules/hearing.ts b/packages/stage-ui/src/stores/modules/hearing.ts
index 9898dbcae..554c7999d 100644
--- a/packages/stage-ui/src/stores/modules/hearing.ts
+++ b/packages/stage-ui/src/stores/modules/hearing.ts
@@ -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
}
+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 = {
'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('')
const autoSendEnabled = useLocalStorageManualReset('settings/hearing/auto-send-enabled', false)
const autoSendDelay = useLocalStorageManualReset('settings/hearing/auto-send-delay', 2000) // Default 2 seconds
+ const confidenceThreshold = useLocalStorageManualReset('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,
diff --git a/packages/ui/src/components/form/field/field-range.vue b/packages/ui/src/components/form/field/field-range.vue
index bc6776328..b9e1bd6aa 100644
--- a/packages/ui/src/components/form/field/field-range.vue
+++ b/packages/ui/src/components/form/field/field-range.vue
@@ -36,9 +36,9 @@ const modelValue = defineModel({ required: true })