[Hearing] Confidence filter is enabled but the provider did

Error message

[Hearing] Confidence filter is enabled but the provider did not return verbose_json segments. Filtering has no effect.

What it means

With the hearing confidence filter enabled (confidenceThreshold above the disabled sentinel), the module requests responseFormat 'verbose_json' from the transcription provider. Some OpenAI-compatible endpoints accept verbose_json but return a payload without a segments array; when response.segments is missing, the confidence filter cannot run, verboseJsonNotSupported is flagged for the UI, and the unfiltered normalizeGeneratedTranscriptionText fallback text is returned instead.

Source

Thrown at packages/stage-ui/src/stores/modules/hearing.ts:523

        file: normalizedInput.file,
        fileName: resolveTranscriptionFileName(normalizedInput.file, normalizedInput.fileName),
        responseFormat: useVerboseJson ? 'verbose_json' : format,
      })

      if (useVerboseJson) {
        if (response.segments) {
          verboseJsonNotSupported.value = false
          const filteredText = filterTranscriptionByConfidence(response.segments, confidenceThreshold.value)
          emitSucceeded(filteredText.length, false)
          return {
            mode: 'generate',
            ...response,
            text: filteredText,
          }
        }
        else {
          verboseJsonNotSupported.value = true
          console.warn('[Hearing] Confidence filter is enabled but the provider did not return verbose_json segments. Filtering has no effect.')
        }
      }

      const fallbackText = normalizeGeneratedTranscriptionText(response)
      emitSucceeded(fallbackText.length, false)
      return {
        mode: 'generate',
        ...response,
        text: fallbackText,
      }
    }
    catch (err) {
      emitFailed(err)
      throw err
    }
  }

  return {

View on GitHub (pinned to f679616c34)

Solutions

Use a provider that honors response_format=verbose_json (e.g. the OpenAI Whisper API) when confidence filtering matters
Upgrade the local whisper server to a build that emits segments in verbose_json
Disable the confidence threshold in hearing settings - useVerboseJson then becomes false and the warning stops
Surface verboseJsonNotSupported in the UI so users see the filter is inert instead of only a console warn

Example fix

// before: endpoint ignores verbose_json, segments undefined
const response = await generateTranscription({ ...request, responseFormat: 'verbose_json' })
filterTranscriptionByConfidence(response.segments ?? [], threshold)

// after: capability-check before filtering
if (useVerboseJson && Array.isArray(response.segments)) {
  return { ...response, text: filterTranscriptionByConfidence(response.segments, threshold) }
}
verboseJsonNotSupported.value = true
return { ...response, text: normalizeGeneratedTranscriptionText(response) }

Defensive patterns

Strategy: validation

Validate before calling

const probe = await generateTranscription({ ...request, responseFormat: 'verbose_json' })
if (!Array.isArray(probe.segments)) {
  verboseJsonNotSupported.value = true
  // disable the confidence filter for this provider
}

Type guard

interface SegmentedTranscription {
  segments?: Array<{ text: string, avg_logprob?: number }>
}
function hasVerboseSegments(r: SegmentedTranscription): r is SegmentedTranscription & { segments: NonNullable<SegmentedTranscription['segments']> } {
  return Array.isArray(r.segments) && r.segments.length > 0
}

Prevention

Probe provider verbose_json support once at setup and cache the result (verboseJsonNotSupported)
Surface filter inactivity in the UI instead of relying on the console warn
Keep the normalizeGeneratedTranscriptionText fallback wired so text is never lost

When it happens

Trigger: Using an OpenAI-compatible provider whose /audio/transcriptions endpoint ignores response_format=verbose_json (local whisper.cpp / faster-whisper shims, gateways that rewrite responses); providers that omit segments for very short clips; any generateTranscription response lacking segments while useVerboseJson is true.

Common situations: Self-hosted whisper servers with partial OpenAI API compatibility; third-party proxies normalizing verbose_json to plain text; provider upgrades dropping segments support; confidence slider enabled against such a provider.

Related errors

AI-assisted analysis of moeru-ai/airi@f679616c34 (2026-08-18). Data as JSON: /api/errors/c82baa36cc473f13. Report an issue: GitHub.