TotalView Logo

Analysis & Scoring

Processing pipeline and score calculation for call analysis

#Analysis & Scoring

#Overview

After you upload a call, TotalView AI processes it through a multi-stage pipeline that converts audio to actionable insights. This document explains how that pipeline works and how scores are calculated.

#Processing Pipeline

Every call goes through these stages:

Upload → PENDING → TRANSCRIBING → ANALYZING → COMPLETED
              ↓           ↓            ↓
           FAILED      FAILED       FAILED

#Stage 1: Pending

What happens: The call is queued for processing.

Duration: Seconds to minutes (depends on queue size)

Status: PENDING

The call waits in a processing queue. During busy periods, there may be a short delay before transcription begins.

#Stage 2: Transcribing

What happens: Audio is converted to text using AI speech recognition.

Duration: Typically 1-5 minutes (proportional to call length)

Status: TRANSCRIBING

The transcription service:

  1. Analyzes the audio stream
  2. Identifies individual speakers (diarization)
  3. Converts speech to text
  4. Assigns speaker labels (Speaker A, Speaker B)
  5. Timestamps each segment

#Stage 3: Analyzing

What happens: AI evaluates the transcript against each quality check.

Duration: Typically 1-3 minutes

Status: ANALYZING

For each quality check, the AI:

  1. Reads the transcript
  2. Applies the evaluation prompt
  3. Determines pass/fail/N/A
  4. Assigns a score (0-100)
  5. Extracts evidence quotes
  6. Records reasoning

Additional analysis includes:

  • Overall sentiment detection
  • Executive summary generation
  • Key topic extraction

#Stage 4: Complete

What happens: All processing finished successfully.

Status: COMPLETED

The call is now ready for review with:

  • Full transcript with speaker labels
  • Check results with scores
  • AI-generated summary
  • Sentiment analysis
  • Overall quality score

#Failed Status

If any stage encounters an error:

Status: FAILED

Common failure reasons:

  • Audio quality too poor for transcription
  • Transcription service timeout
  • AI analysis error
  • Rate limit exceeded

Failed calls can be retried from the call detail page.

#Transcription

#How It Works

TotalView AI uses two transcription providers:

Provider Style Speed Cost Best For
AssemblyAI Async (webhook) Moderate Higher Maximum accuracy
Deepgram Sync Faster Lower High volume

Your campaign settings determine which provider is used.

#Speaker Diarization

The transcription identifies different speakers:

  • Speaker A: First speaker detected
  • Speaker B: Second speaker detected
  • Additional speakers: Speaker C, D, etc.

#Agent Detection

TotalView AI automatically detects which speaker is the agent:

  1. Analyzes speaking patterns
  2. Looks for introduction phrases ("Hi, this is John from...")
  3. Assigns the agent_speaker_label field

This helps quality checks focus on agent behavior.

#Transcript Structure

Each transcript contains segments:

Field Description
Speaker Label Who is speaking (A, B, etc.)
Text What was said
Start Time When segment begins (milliseconds)
End Time When segment ends (milliseconds)
Confidence Transcription confidence (0-1)

#Quality Check Analysis

#How Checks Are Evaluated

For each check in the campaign:

  1. Context Provided: The AI receives:

    • Full transcript
    • Check name and type
    • Evaluation prompt
    • Speaker information

  2. Evaluation: The AI determines:

    • Does this check apply? (Yes/No/N/A)
    • If yes, was the criteria met?
    • What evidence supports this conclusion?

  3. Scoring: Based on how well criteria were met:

    • 100: Perfectly met
    • 70-99: Met with minor issues
    • 40-69: Partially met
    • 1-39: Mostly not met
    • 0: Completely failed

  4. Result: The check produces:

    • Result type (pass/fail/na)
    • Score (0-100)
    • Reasoning (explanation)
    • Evidence (transcript quotes)

#Check Priority

Checks are evaluated in priority order (highest first). This allows:

  • Critical compliance checks to run first
  • Context from early checks to inform later ones

#Score Calculation

#Individual Check Scores

Each check receives a score from 0 to 100:

Score Range Interpretation Result
70-100 Criteria met Pass
0-69 Criteria not met Fail
N/A Not applicable Excluded

#Overall Quality Score

The overall score is a weighted average of all check scores:

Overall Score = Σ(Check Score × Weight) ÷ Σ(Weights)

#Example Calculation

Check Score Weight Contribution
Opening Greeting 95 1.0 95
Customer Verification 100 2.0 200
Issue Resolution 75 1.5 112.5
Closing 80 1.0 80
Totals 5.5 487.5

Overall Score = 487.5 ÷ 5.5 = 88.6

#N/A Checks

Checks marked as N/A are excluded from the overall score calculation:

  • They don't add to the numerator
  • They don't add to the denominator (weight total)

This prevents irrelevant checks from affecting the score.

#Score Thresholds

Default thresholds (configurable per campaign):

Threshold Default Meaning
Pass/Fail 70 Checks ≥70 pass
Flag for Review Configurable Calls below this are flagged

#Additional AI Analysis

Beyond quality checks, each call receives:

#Executive Summary

A 2-3 sentence overview of the call:

  • What was the call about?
  • What was the outcome?
  • Any notable issues?

#Sentiment Analysis

Overall emotional tone of the call:

Sentiment Meaning
POSITIVE Customer expressed satisfaction
NEUTRAL No strong positive or negative
NEGATIVE Customer expressed frustration

Sentiment includes reasoning explaining the classification.

#Key Topics

Main subjects discussed in the call:

  • Product/service names
  • Issue categories
  • Customer concerns

#Processing Times

Typical end-to-end processing times:

Call Duration Transcription Analysis Total
< 5 minutes 1-2 min 1 min 2-3 min
5-15 minutes 2-3 min 1-2 min 3-5 min
15-30 minutes 3-5 min 2-3 min 5-8 min
30-60 minutes 5-10 min 3-5 min 8-15 min

Factors affecting processing time:

  • Audio quality (poor quality = slower)
  • Number of checks (more checks = longer analysis)
  • System load (peak times may be slower)
  • Provider (Deepgram faster than AssemblyAI)

#Re-Running Analysis

You can re-run analysis on previously processed calls:

#When to Re-Run

  • Added or modified quality checks
  • Want to apply new checks to old calls
  • Previous analysis had issues

#How to Re-Run

Single Call:

  1. Open the call detail page
  2. Click Re-analyze
  3. Confirm the action
  4. Wait for processing to complete

Bulk Re-Analysis:

  1. Navigate to campaign Calls page
  2. Select multiple calls using checkboxes
  3. Click Bulk Actions → Re-analyze
  4. Monitor progress in the Jobs page

#What Happens

  • New analysis replaces the previous one
  • Old check results are preserved with version numbers
  • Score history is maintained
  • Transcript remains unchanged

#Version History

TotalView AI tracks analysis versions:

Version Created When
v1 Original analysis
v2 First re-analysis
v3+ Subsequent re-analyses

You can view historical check results to see how scores changed.

#Processing Costs

Processing calls consumes API credits for:

#Transcription

Provider Billing
AssemblyAI Per audio minute
Deepgram Per audio second (cheaper)

#AI Analysis

Component Billing
Check evaluation Per token (input + output)
Summary generation Per token
Sentiment analysis Included in analysis

#Viewing Usage

  1. Go to Settings → Usage (or dashboard sidebar)
  2. View usage by:
    • Service (transcription, analysis)
    • Time period
    • Campaign

#Troubleshooting

#Call Stuck on Pending

Possible causes:

  • Processing queue backlog
  • Auto-processing disabled

Solutions:

  1. Wait a few minutes
  2. Check if auto-processing is enabled in campaign settings
  3. Manually trigger from call detail page

#Call Stuck on Transcribing

Possible causes:

  • Long audio file
  • Transcription service issues

Solutions:

  1. Wait longer for long files (30+ min = 10+ min processing)
  2. Check Jobs page for status
  3. If stuck >30 min, retry

#Analysis Failed

Possible causes:

  • Rate limit exceeded
  • AI service error
  • Invalid transcript

Solutions:

  1. Click Retry on the call detail page
  2. Check error message for details
  3. Contact support if persistent

Learn More: Troubleshooting for detailed solutions.

#FAQ

#How is the overall score calculated?

It's a weighted average of all check scores. Checks with higher weights have more influence. N/A checks are excluded from the calculation.

#Can I change how the score is calculated?

You can adjust individual check weights to change their influence. The weighted average formula itself is fixed.

#What if all checks are N/A?

If every check returns N/A, the call will have no overall score (null). This typically indicates the transcript was too short or the checks didn't match the call content.

#Does re-analysis change the transcript?

No. Re-analysis only re-runs the quality check evaluation. To get a new transcript, delete and re-upload the call.

#Why did my check score 50 instead of 0 or 100?

The AI uses a 0-100 scale to indicate degree of compliance. A score of 50 might mean the criteria was partially met — the agent did some but not all of what was expected.

#How do I know which speaker is the agent?

The system auto-detects the agent based on introduction patterns. You can see the assigned speaker label in the transcript view. If incorrect, quality checks still evaluate based on the full conversation.