> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cekura.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Pre-Defined Metrics
> Standard metrics for evaluating voice agent performance across accuracy, conversation quality, customer experience, and speech quality
export const CopyPageButton = () => {
if (typeof window !== 'undefined') {
setTimeout(function () {
if (document.getElementById('ck-tools')) return;
var anchor = document.getElementById('content-area') || document.querySelector('.mdx-content');
if (!anchor) return;
if (!document.getElementById('ck-style')) {
var s = document.createElement('style');
s.id = 'ck-style';
s.textContent = '#ck-tools{position:absolute;top:6px;right:0;z-index:100;font-family:inherit;}' + '.ck-row{display:inline-flex;align-items:stretch;border:1px solid rgba(0,0,0,0.15);border-radius:8px;overflow:hidden;background:#fff;}' + ':root.dark .ck-row{background:rgba(255,255,255,0.06);border-color:rgba(255,255,255,0.12);}' + '.ck-btn{padding:5px 12px;border:none;background:none;cursor:pointer;font-size:13px;font-weight:500;font-family:inherit;color:#374151;}' + ':root.dark .ck-btn{color:#d1d5db;}' + '.ck-btn:hover{background:rgba(0,0,0,0.04);}' + ':root.dark .ck-btn:hover{background:rgba(255,255,255,0.06);}' + '.ck-chevron{padding:5px 8px;border:none;background:none;cursor:pointer;font-size:14px;font-family:inherit;color:#374151;}' + ':root.dark .ck-chevron{color:#d1d5db;}' + '.ck-chevron:hover{background:rgba(0,0,0,0.04);}' + ':root.dark .ck-chevron:hover{background:rgba(255,255,255,0.06);}' + '.ck-divider{width:1px;background:rgba(0,0,0,0.12);flex-shrink:0;}' + ':root.dark .ck-divider{background:rgba(255,255,255,0.12);}' + '.ck-dd{position:absolute;top:calc(100% + 4px);right:0;min-width:180px;background:#fff;border:1px solid rgba(0,0,0,0.12);border-radius:8px;box-shadow:0 4px 12px rgba(0,0,0,0.1);padding:4px;display:none;z-index:200;}' + ':root.dark .ck-dd{background:#1f2937;border-color:rgba(255,255,255,0.1);box-shadow:0 4px 16px rgba(0,0,0,0.35);}' + '.ck-item{display:block;width:100%;padding:7px 12px;border:none;background:none;border-radius:6px;cursor:pointer;font-size:13px;font-family:inherit;text-align:left;color:#374151;}' + ':root.dark .ck-item{color:#d1d5db;}' + '.ck-item:hover{background:rgba(0,0,0,0.05);}' + ':root.dark .ck-item:hover{background:rgba(255,255,255,0.07);}';
document.head.appendChild(s);
}
var wrap = document.createElement('div');
wrap.id = 'ck-tools';
var row = document.createElement('div');
row.className = 'ck-row';
var mainBtn = document.createElement('button');
mainBtn.className = 'ck-btn';
mainBtn.textContent = 'Copy page';
var divider = document.createElement('span');
divider.className = 'ck-divider';
var chevron = document.createElement('button');
chevron.className = 'ck-chevron';
chevron.textContent = '▾';
var dd = document.createElement('div');
dd.className = 'ck-dd';
function closeDD() {
dd.style.display = 'none';
}
function openDD() {
dd.style.display = 'block';
}
chevron.onclick = function (e) {
e.stopPropagation();
if (dd.style.display === 'block') {
closeDD();
} else {
openDD();
}
};
document.addEventListener('click', function (e) {
if (!e.target.closest('#ck-tools')) {
closeDD();
}
});
document.addEventListener('keydown', function (e) {
if (e.key === 'Escape') {
closeDD();
}
});
function makeItem(label, fn) {
var b = document.createElement('button');
b.className = 'ck-item';
b.textContent = label;
b.onclick = function () {
fn();
closeDD();
};
return b;
}
function getMarkdown() {
var walk = function (node) {
if (!node) return '';
if (node.nodeType === 3) return node.textContent || '';
if (node.nodeType !== 1) return '';
var tag = node.tagName.toLowerCase();
var skip = ['script', 'style', 'svg', 'noscript', 'button', 'iframe'];
if (skip.indexOf(tag) !== -1) return '';
if (node.id === 'ck-tools') return '';
var ch = Array.from(node.childNodes).map(walk).join('');
if (tag === 'h1') return '\n# ' + ch.trim() + '\n\n';
if (tag === 'h2') return '\n## ' + ch.trim() + '\n\n';
if (tag === 'h3') return '\n### ' + ch.trim() + '\n\n';
if (tag === 'p') return '\n' + ch.trim() + '\n\n';
if (tag === 'pre') return '\n```\n' + node.textContent.trim() + '\n```\n\n';
if (tag === 'li') return '- ' + ch.trim() + '\n';
if (tag === 'code') return '`' + ch.trim() + '`';
return ch;
};
var content = document.querySelector('.mdx-content') || document.getElementById('content-area') || document.body;
return walk(content).replace(/\n\n\n+/g, '\n\n').trim();
}
function copyMd() {
var md = getMarkdown();
navigator.clipboard.writeText(md).then(function () {
mainBtn.textContent = 'Copied!';
setTimeout(function () {
mainBtn.textContent = 'Copy page';
}, 2000);
});
}
function viewMd() {
var md = getMarkdown();
var safe = md.split('&').join('&').split('<').join('<').split('>').join('>');
var html = '' + safe + '';
window.open(URL.createObjectURL(new Blob([html], {
type: 'text/html'
})), '_blank');
}
function openClaude() {
var prompt = 'Can you read this Cekura docs page ' + window.location.href + ' so I can ask you questions?';
window.open('https://claude.ai/new?q=' + encodeURIComponent(prompt), '_blank');
}
mainBtn.onclick = copyMd;
dd.appendChild(makeItem('Copy page', copyMd));
dd.appendChild(makeItem('View as Markdown', viewMd));
dd.appendChild(makeItem('Open in Claude', openClaude));
row.appendChild(mainBtn);
row.appendChild(divider);
row.appendChild(chevron);
wrap.appendChild(row);
wrap.appendChild(dd);
anchor.style.position = 'relative';
anchor.insertBefore(wrap, anchor.firstChild);
}, 50);
}
return null;
};
These are standard metrics applicable across domains defined by Cekura. They are organized into four categories based on what aspect of your voice agent they evaluate.
***
## Accuracy Metrics
These metrics evaluate whether your agent provides correct and consistent information.
**Result**: Pass / Review Required / Failed | **Cost**: 0 credits
Evaluates whether the Main Agent achieved the goal specified in the evaluator's expected outcome prompt. An LLM analyzes the conversation to determine if the intended outcome was reached.
**Requirements**: Set `expected_outcome_prompt` in your evaluator configuration describing what success looks like.
**Interpretation**:
* **Pass**: Main Agent achieved the expected outcome
* **Review Required**: Outcome unclear, manual review recommended
* **Failed**: Main Agent did not achieve the expected outcome
**Result**: True/False | **Cost**: 0.6 credits per call
Detects when the Main Agent provides information that contradicts or isn't supported by the uploaded Knowledge Base files. A Knowledge Base is a collection of files uploaded to the agent containing reference information for fact-checking. An LLM compares Main Agent responses against the Knowledge Base content.
**Requirements**: Upload Knowledge Base files to your agent containing the source of truth.
**Interpretation**:
* **True**: No hallucinations detected (Main Agent stayed factual)
* **False**: Main Agent provided unsupported or contradictory information
**Result**: True/False | **Cost**: 0.2 credits per call
Evaluates whether the Main Agent's responses were relevant and appropriate to the conversation context. An LLM analyzes each response to determine if it addressed what the Testing Agent was asking.
**Interpretation**:
* **True**: Responses were relevant and on-topic
* **False**: Main Agent gave off-topic or inappropriate responses
**Result**: True/False | **Cost**: 0.2 credits per call
Detects inconsistencies in the Main Agent's responses during a call. An LLM checks for two specific issues:
1. Testing Agent provides information (e.g., their name) and the Main Agent repeats it back incorrectly
2. Main Agent makes contradictory statements (e.g., says one thing early in the call, then contradicts it later)
**Interpretation**:
* **True**: Main Agent maintained consistent information throughout
* **False**: Inconsistencies or contradictions detected
**Result**: True/False | **Cost**: 0 credits
Checks whether any tool calls made by the Main Agent resulted in errors.
**Interpretation**:
* **True**: All tool calls succeeded
* **False**: One or more tool calls returned an error
**Result**: Score (0-100) | **Cost**: 0 credits (Runs) / 1.0 credits per minute (Call Logs)
Evaluates speech-to-text accuracy differently depending on the context:
**For Call Logs**: Uses two separate state-of-the-art transcription models to generate ground truth transcriptions. Compares these against the candidate transcript to find inconsistencies, with the score based on the number of errors found. No text normalisation is applied before scoring — the raw transcript is sent directly to the LLM judge so it can assess exactly what was transcribed.
**For Runs/Simulations**: Measures how accurately the provider's speech-to-text transcribed the **testing agent** (the simulated caller). It compares the provider transcript against Cekura's ground-truth transcript of what the testing agent actually said — Cekura generates the testing agent's speech, so it knows the exact words spoken. Only the testing agent's turns are scored: Cekura has no independent ground truth for the main agent's turns (both transcripts there are just speech-to-text of the same audio), so scoring them would be circular. Use this to catch cases where the provider's transcription mishears the caller — a wrong word, a dropped digit in a number, or a name transcribed incorrectly.
**What is Word Error Rate (WER)?** WER is the standard speech-recognition accuracy measure — the fraction of spoken words the speech-to-text got wrong:
```
WER = (substitutions + deletions + insertions) / words in reference
```
A WER of 0% means a perfect transcription; lower is better. The explanation always reports the WER percentage for the run.
**How the score is calculated.** The score is *not* the raw WER. Each transcription error is weighted by how much the mistranscribed word matters, and the score comes from the total **weighted error count**:
| Word type | Weight per error |
| ------------------------------------------------------------ | ---------------------------------------------------------------------------- |
| Names, common nouns, numbers | **1.0** |
| Negation flips (e.g. "can" vs "can't", a dropped "no"/"not") | **1.0** |
| Verbs | **0.5** |
| Everything else (articles, pronouns, fillers, prepositions…) | **0** toward the weighted count — treated as a *minor variation* (see below) |
The same distinct significant word mistranscribed multiple times is only counted once. This POS-weighted approach works across common languages including Spanish. The weighted error count maps to a base score:
| Base score | Weighted errors |
| ---------- | --------------- |
| 100% | 0 |
| 80% | 1–3 |
| 60% | 4–6 |
| 40% | 7–12 |
| 20% | 13+ |
**Minor variations.** Every remaining difference — articles, pronouns, fillers, prepositions and other function words — is a *minor variation*. These don't count toward the weighted error count, but each one applies a small **absolute 0.1% penalty** to the score. So a transcript with no significant errors and, say, 3 minor variations scores 99.7% instead of a flat 100%, and a call scored 60% on significant errors with 6 minor variations lands at 59.4%. Significant errors still dominate; the minor penalty only lets you tell a flawless transcript apart from one with a few small differences by looking at the score alone. This applies to both Runs and Call Logs.
**Rendering-only differences are not errors.** Differences that only change how a value is *rendered* are treated as equivalent and don't count at all — spelled-out vs digit numbers, and spelled-out vs abbreviated units (e.g. "five kilowatts" vs "5 kW", "two hundred thirty volt" vs "230 V"). Only a genuine change in the value or the unit ("20 watts" vs "20 volts") is a real error. Backchannel acknowledgements ("uh-huh" vs "aha") are treated as minor variations.
**Reading the explanation.** Errors are grouped into **Major Variations** (drive the base score) and **Minor Variations** (each applies the 0.1% penalty), so you can see exactly which mistranscriptions drove the score. Both sections always appear — `None` means there were no errors of that kind. A summary at the bottom shows the **Metric Score** and the **WER** percentage.
**Interpretation**: Higher scores indicate better transcription accuracy. Use this metric when you need to verify that the agent's TTS is rendering content correctly (correct words, correct numbers), not just that speech is audible.
**Result**: True/False | **Cost**: 0.2 credits per call
Detects whether the call reached a voicemail system instead of a live person. An LLM analyzes the transcript for voicemail indicators (greeting messages, beeps, recording prompts).
**Interpretation**:
* **True**: Call reached voicemail
* **False**: Call connected to a live person
***
## Conversation Quality Metrics
These metrics evaluate the flow and dynamics of the conversation.
**Result**: Count (number of interruptions) | **Cost**: 0 credits
Counts how many times the Main Agent started speaking while the Testing Agent was still talking. An interruption is when one speaker begins talking while the other is still speaking. Uses Voice Activity Detection (VAD) to find the timestamps for each turn of each speaker and precisely detect overlapping speech.
**Requirements**: Stereo audio recording with separate channels for each speaker.
**Interpretation**: Lower is better. Frequent interruptions indicate the Main Agent isn't properly waiting for the Testing Agent to finish speaking.
**Result**: Numeric (milliseconds) | **Cost**: 0 credits
Measures how long it takes for the Main Agent to stop speaking after the Testing Agent interrupts. Uses VAD to find the timestamps for each turn and detect when the Testing Agent starts speaking over the Main Agent.
**Requirements**: Stereo audio recording with separate channels for each speaker.
**Interpretation**: Lower is better. A responsive Main Agent should stop quickly (under 500ms) when interrupted.
**Result**: Count (number of interruptions) | **Cost**: 0 credits
Counts how many times the Testing Agent started speaking while the Main Agent was still talking. Uses VAD to find the timestamps for each turn and detect overlapping speech.
**Requirements**: Stereo audio recording with separate channels for each speaker.
**Interpretation**: High counts may indicate Testing Agent frustration, Main Agent speaking too long, or poor turn-taking.
**Result**: Score (0-5, shown as a percentage on the Results page) | **Cost**: 0 credits
Normalizes how often the Main Agent interrupts the Testing Agent against the number of turns: `5 * (1 - interruptions / turns)`, clamped to 0-5. Uses VAD on stereo audio where available, otherwise transcript analysis. The Results page shows it as a percentage of the 5-point maximum (e.g. 5.0 → 100%).
**Interpretation**: Higher is better. 5 (100%) means no interruptions; lower scores indicate the Main Agent increasingly talks over the Testing Agent.
**Result**: Numeric (average milliseconds) | **Cost**: 0 credits
Measures the response time between the Testing Agent finishing speaking and the Main Agent starting its response. Latency is calculated for every Main Agent turn throughout the call. With stereo audio, VAD is used to find precise timestamps for each speaker turn.
We compute and display percentile statistics: P25, P50, P75, P90, P95, and P99.
**Interpretation**: Lower is better. Latency under 2000ms is generally good.
**Result**: Count (number of repetitions) | **Cost**: 0.2 credits per call
Identifies instances where the Main Agent unnecessarily repeated information it had already provided. An LLM analyzes the conversation for redundant statements.
**Interpretation**: Lower is better. Repetition wastes time and can frustrate the Testing Agent.
**Result**: Score (0-5) | **Cost**: 0.2 credits per call
The 0-5 score companion to **Unnecessary Repetition Count**, normalizing repetitions against the number of turns: `5 * (1 - repetitions / turns)`, clamped to 0-5. An LLM flags when the Main Agent re-confirms the same, unchanged information two or more times.
**Interpretation**: Higher is better. 5 means no unnecessary repetition; lower scores mean the agent keeps re-confirming the same information.
**Result**: Score (1-5) | **Cost**: 0.2 credits per call
Evaluates how appropriately concise the Main Agent is across the conversation. An LLM classifies each Main Agent turn as Concise, Padded (filler, restated questions, repeated information, disclaimer spam), or Overloaded (more information, options, or follow-up questions than the user asked for), then aggregates to a 1-5 score. Necessary detail the user explicitly asked for is not penalised.
**Interpretation**:
* **5**: Consistently concise and well-calibrated to user intent
* **4**: Mostly concise; a few overlong turns
* **3**: Mixed; several turns are too long
* **2**: Verbosity is the dominant issue across most turns
* **1**: Agent is bloated throughout
**Result**: True/False | **Cost**: 0 credits
Detects prolonged silence periods where both the Main Agent and Testing Agent are silent, which may indicate technical issues or agent problems.
**Requirements**: Audio recording.
**Configuration**: Set `silence_duration` in the metric configuration (default: 10 seconds).
**Interpretation**:
* **True**: No problematic silence detected
* **False**: Extended mutual silence exceeding the threshold was detected
**Result**: True/False | **Cost**: 0 credits
Detects when the Main Agent fails to respond within the configured timeout after the Testing Agent finishes their turn, indicating potential infrastructure or connectivity problems.
**Requirements**: Audio recording.
**Configuration**: Set `infra_issues_timeout` in the metric configuration (default: 10 seconds).
**Interpretation**:
* **True**: No infrastructure issues detected
* **False**: Main Agent failed to respond within the timeout after the Testing Agent finished speaking
**Result**: True/False | **Cost**: 0.2 credits per call
Evaluates whether the Main Agent ended the call appropriately. An LLM analyzes whether the Main Agent wrapped up the conversation properly before ending.
**Interpretation**:
* **True**: Call was ended appropriately by the Main Agent
* **False**: Main Agent ended call abruptly or inappropriately
**Result**: True/False | **Cost**: 0.2 credits per call
Evaluates whether the Testing Agent ended the call early, which may indicate poor experience or unresolved issues. An LLM analyzes the conversation to determine if the call ended prematurely.
**Interpretation**:
* **True**: Call ended at a natural conclusion point
* **False**: Testing Agent ended call early, suggesting dissatisfaction
***
## Customer Experience Metrics
These metrics evaluate the Testing Agent's experience and satisfaction with the conversation.
**Result**: Score (0-100) | **Cost**: 0.2 credits per call
Evaluates overall customer satisfaction based on two dimensions:
**1. Customer Sentiment** - Evaluates the Testing Agent's (customer's) tone throughout the call:
* Positive: Clear expressions of gratitude like "Thank you so much for your help" = 5 points
* Neutral: Simple thanks, cooperative tone, matter-of-fact responses = 5 points
* Negative: Explicit frustration, harsh language, complaints = 1 point
**2. Cooperation** - Evaluates whether the Main Agent helped the Testing Agent:
* Fully cooperative / No issues = 5 points
* Somewhat uncooperative = 3 points
* Refused to help / Obstructed = 1 point
The final score is the average of both dimensions, scaled to 0-100.
**Interpretation**: Higher is better. Scores above 70 indicate good customer satisfaction.
**Result**: Enum (one of your configured stages) | **Cost**: 0.2 credits per call
Identifies at which stage of the conversation the call dropped or ended. An LLM maps the conversation endpoint to one of your predefined stages.
**Requirements**: Configure `dropoff_nodes` on your agent with the conversation stages you want to track (e.g., "greeting", "information\_gathering", "resolution", "closing").
**Interpretation**: Helps identify where in your conversation flow Testing Agents are dropping off, enabling targeted improvements.
**Result**: Enum (positive / neutral / negative) | **Cost**: 0.2 credits per call
Determines the Testing Agent's overall sentiment toward the Main Agent based on the conversation transcript. An LLM analyzes the Testing Agent's language, tone, and responses.
**Classification criteria**:
* **Positive**: Only when the Testing Agent is clearly very grateful with phrases like "Thank you so much for your help", "I really appreciate this", "You've been so helpful"
* **Negative**: Explicit frustration, harsh language, complaints, or aggressive tone
* **Neutral**: Simple "thanks", cooperative tone, matter-of-fact responses, or when sentiment is unclear
**Note**: Neutral is the default value when sentiment is uncertain.
**Interpretation**:
* **Positive**: Testing Agent seemed very satisfied or grateful
* **Neutral**: Testing Agent showed no strong emotion
* **Negative**: Testing Agent seemed frustrated or dissatisfied
**Result**: Enum (one of your configured topics) | **Cost**: 0.2 credits per call
Categorizes the call into one of your predefined topics. An LLM analyzes the conversation to determine the primary subject matter.
**Requirements**: Configure `topic_nodes` on your agent with the topics you want to track (e.g., "billing", "technical\_support", "sales", "general\_inquiry").
**Interpretation**: Helps understand call volume distribution across different topics for resource planning and analysis.
***
## Speech Quality Metrics
These metrics evaluate the audio and speech characteristics of the Main Agent.
**Result**: Numeric (Hertz) | **Cost**: 0 credits
Measures the average pitch frequency of the Main Agent's voice during the call using pitch extraction algorithms.
**Requirements**: Audio recording.
**Result**: True/False | **Cost**: 0.2 credits per minute
Detects nonsensical or garbled speech from the Main Agent. A multimodal LLM analyzes the Main Agent's audio to identify unintelligible segments, nonsense sounds, or garbled speech.
**Requirements**: Stereo audio recording with separate channels.
**Interpretation**:
* **True**: Speech was clear and intelligible
* **False**: Gibberish or garbled speech detected
**Result**: True/False | **Cost**: 0.2 credits per call
Checks whether certain words were spelled out letter-by-letter correctly in the audio (e.g., spelling out "J-O-H-N" for a name). A multimodal LLM analyzes the Main Agent's audio to verify spelling.
**Requirements**: Audio recording. Configure `spelling_word_types` on your agent specifying which types of words should be spelled out (e.g., "name", "email", "confirmation\_code").
**Interpretation**:
* **True**: Every instance of every word of the selected category was correctly spelled out in the audio
* **False**: Spelling errors detected or words not spelled out when required
**Result**: Score (0-100) | **Cost**: 0.2 credits per call
Evaluates pronunciation accuracy for specific words you define.
**Requirements**: Audio recording. Configure `pronunciation_words` on your agent as a list of word-phoneme pairs (e.g., `[["Cekura", "suh-KYUR-uh"]]`).
**Interpretation**: Higher scores indicate better pronunciation accuracy. Useful for brand names or technical terms.
**Result**: True/False | **Cost**: 0.2 credits per call
Detects abrupt or unnatural changes in the Main Agent's speaking rate during the call using an ML model.
**Requirements**: Audio recording. Currently supports English only.
**Interpretation**:
* **True**: Speaking rate was consistent and natural
* **False**: Unnatural speaking rate changes detected
**Result**: Numeric (0.0 to 1.0) | **Cost**: 0 credits
Calculates the ratio of Main Agent speaking time to total call duration. Uses VAD to find the timestamps for each turn of each speaker for accurate speaker separation.
**Requirements**: Stereo audio recording with separate channels.
**Interpretation**: A ratio around 0.4-0.6 is typical. Very high ratios may indicate the Main Agent is dominating the conversation; very low ratios may indicate the Main Agent isn't being helpful enough.
**Result**: True/False | **Cost**: 0.2 credits per call
Detects unexpected speaker changes during the Main Agent's speaking turns using an ML model for voice analysis.
**Requirements**: Audio recording.
**Interpretation**:
* **True**: Consistent speaker throughout Main Agent turns
* **False**: Unexpected voice change detected (may indicate system issues)
**Result**: Score (0-100) | **Cost**: 0.2 credits per call
Evaluates the overall voice quality of the Main Agent's audio using an ML model. Specifically analyzes clarity (how clear and understandable the voice is) and jitter (variations in audio timing that can affect quality) on the Main Agent's audio channel.
**Requirements**: Audio recording.
**Example of low voice clarity**:
**Interpretation**: Higher is better. Scores above 70 indicate good voice quality. Low scores may indicate audio issues, background noise, or voice synthesis problems.
**Result**: Numeric (words per minute) | **Cost**: 0 credits
Calculates the Main Agent's speaking speed based on transcript word count and speaking duration.
**Requirements**: Audio recording.