Voice Judge

Retrieve audio-native voice quality scores for your voice agent calls.

The Voice Judge evaluates voice agent calls directly from audio recordings, scoring each call across 10 quality dimensions. Scores measure the voice experience - pronunciation, clarity, pacing, interruption handling - independent of conversational logic or agent prompting.

Voice judge results are produced by a scheduled batch evaluation job. Once scored, results are available per-service through the API.

Endpoints

List Recent Voice Judge Results

Latency: 500ms-2s. This endpoint reads from the analytics warehouse, not the primary database.

GET /v1/{workspace_id}/services/{service_id}/voice-judge/recent

Returns the most recent per-call voice quality scores for a service, ordered newest first.

Path Parameters

Parameter	Type	Description
workspace_id	string (UUID)	Workspace identifier
service_id	string (UUID)	Service identifier

Query Parameters

Parameter	Type	Default	Description
limit	integer	20	Max rows to return. Min 1, max 100.

Response

{
  "service_id": "string",
  "count": 2,
  "items": [
    {
      "call_sid": "string",
      "call_entity_id": "string | null",
      "service_id": "string | null",
      "latency_dead_air_score": 0.9,
      "pronunciation_score": 1.0,
      "clarity_score": 0.25,
      "filler_silence_score": 0.5,
      "interruption_handling_score": 1.0,
      "audio_consistency_score": 1.0,
      "pacing_score": 1.0,
      "warmth_tone_score": 1.0,
      "accent_quality_score": 1.0,
      "voice_identity_score": 1.0,
      "overall_score": 0.865,
      "critical_count": 0,
      "flag_count": 1,
      "warning_count": 1,
      "judge_json": "string | null",
      "computed_at": "2026-05-15T12:00:00Z"
    }
  ]
}

Response Fields

Field	Type	Description
service_id	string	The service these results belong to
count	integer	Number of items returned
items	array	List of voice judge result rows

Voice Judge Result Row

Field	Type	Description
call_sid	string	Call identifier
call_entity_id	string or null	Entity identifier for the call
service_id	string or null	Service identifier
latency_dead_air_score	number or null	Response latency and dead air score (P0). 0.0-1.0
pronunciation_score	number or null	Pronunciation accuracy score (P0). 0.0-1.0
clarity_score	number or null	Speech clarity and intelligibility score (P0). 0.0-1.0
filler_silence_score	number or null	Filler and silence management score (P1). 0.0-1.0
interruption_handling_score	number or null	Barge-in and recovery score (P1). 0.0-1.0
audio_consistency_score	number or null	Audio consistency score (P1). 0.0-1.0
pacing_score	number or null	Speech rate and pausing score (P2). 0.0-1.0
warmth_tone_score	number or null	Emotional tone appropriateness score (P2). 0.0-1.0
accent_quality_score	number or null	Language and accent match score (P2). 0.0-1.0
voice_identity_score	number or null	Voice consistency across the call score (P2). 0.0-1.0
overall_score	number or null	Composite score (arithmetic mean of dimension scores). 0.0-1.0
critical_count	integer or null	Number of dimensions with Critical severity
flag_count	integer or null	Number of dimensions with Flag severity
warning_count	integer or null	Number of dimensions with Warning severity
judge_json	string or null	Raw judge output with per-dimension evidence quotes and severity. Opaque string for UI drill-in.
computed_at	string (ISO 8601) or null	When the evaluation was computed

Score Interpretation

All dimension scores range from 0.0 to 1.0:

Score Range	Severity	Meaning
0.75 - 1.0	None	Meets the bar
0.5 - 0.74	Warning	Minor quality pattern detected
0.25 - 0.49	Flag	Notable quality issue
0.0 - 0.24	Critical	Significant quality problem

Error Responses

Status	Description
404	Service not found in this workspace
503	Analytics warehouse not configured or transiently unavailable

A 200 response with an empty items list means no calls have been evaluated yet for this service. A 503 response means the analytics infrastructure is temporarily unavailable - retry after a short delay.

Dimensions

The voice judge evaluates 10 dimensions, grouped by priority:

P0 - Critical Quality

• Latency and Dead Air - Response latency between turns. Flags prolonged silence (>3s between turns) and extended processing waits without verbal acknowledgment.

• Pronunciation - Correct pronunciation of medical terms, drug names, dates, numbers, and patient names. Critical on any factual read-back error.

• Clarity - Speech intelligibility and clean audio output. Critical on garbled or unintelligible speech.

P1 - Important Quality

• Filler and Silence Management - Graceful handling of processing pauses. Verbal acknowledgment before a pause, no dead air during the wait, no repeated filler phrases, and filler that matches the result being delivered.

• Interruption Handling - Clean barge-in behavior. Agent stops when the caller speaks, no false triggers on background noise, smooth recovery after being interrupted.

• Audio Consistency - Absence of volume spikes, pitch anomalies, mid-word cutoffs, or inconsistent voice timbre across turns.

P2 - Quality Polish

• Pacing - Conversational speech rate with appropriate pauses between pieces of information and slower delivery for sensitive content.

• Warmth and Tone - Emotional appropriateness matched to the caller's state. Flags flat affect, tonal mismatches, or inappropriate emotional tone.

• Accent and Language Quality - Language and accent match to the caller. Critical on wrong language delivery.

• Voice Identity - Consistent agent voice and persona across the entire call.