Manual QA teams review as little as 1% to 2% of contact center calls, leaving the vast majority of customer interactions unreviewed and exposing organizations to silent compliance violations. The industry standard for QA coverage reflects reviewer bandwidth constraints, not operational necessity. That coverage gap means entire agent cohorts, product lines, and complaint categories go unreviewed indefinitely. Those compliance risks stay invisible until a breach forces the conversation.
Automated QA solves the coverage problem, but it introduces a different risk. Every scoring model, LLM prompt, and coaching flag runs on the transcript. If the transcript is wrong, the automation is wrong, and unlike a human reviewer, it will not flag its own uncertainty. Getting transcription right is not a technical detail, it is the operational ceiling for everything downstream.
Anatomy of an effective QA scorecard template
A contact center QA scorecard defines which behaviors are tracked, how they're weighted, and what constitutes a passing score. A QA rubric provides the grading criteria that tell a reviewer or LLM whether each behavior was present, absent, or partially met. Organizations that attempt to automate scoring before defining both typically end up with inconsistent results, because the LLM interprets ambiguous criteria differently across calls.
The table below shows the operational difference between the two approaches:
| Dimension |
Manual QA (1-2% sampling) |
Automated QA (100% coverage) |
| Call coverage |
1-2% of total volume |
100% of calls |
| Review time per call |
~15 minutes |
Seconds to minutes (transcription processes ~60s per hour of audio, human review limited to flagged exceptions) |
| Coaching opportunity |
Limited by reviewer bandwidth |
Every agent, every call |
| Agent blind spots identified |
Limited to sampled calls |
Every agent, every metric, every call |
| Prerequisite |
Consistent evaluator calibration |
Accurate multi-speaker transcription |
Essential metrics for QA scorecards
Every effective scorecard maps agent behaviors to the business KPIs the operation is accountable for. Core categories include script and disclosure compliance (binary, ties to audit risk), problem resolution quality (feeds FCR), customer sentiment trajectory (flags escalation risk), and soft skills (empathy, acknowledgment, professionalism). AHT (Average Handle Time) is tracked alongside quality metrics to identify call-handling inefficiencies, though high AHT may reflect call complexity rather than agent underperformance. Coaching candidates are identified by QA scores, not AHT alone. Each category carries a numeric weight, and both the weights and the grading criteria must be locked before you attempt automation.
Distinguishing scorecards from rubrics
The scorecard is the what, which is the metrics and their weights. The rubric is the how, which is the specific grading criteria for evaluators or automated scoring models. A rubric that says "agent showed empathy" is not parseable by an LLM. A rubric that says "agent acknowledged customer frustration with a specific phrase or paraphrase within two turns of the customer expressing the complaint" can be evaluated consistently. For binary compliance criteria, the rubric is straightforward. For soft-skill criteria, build scoring tiers with anchored examples at each level, and use those examples as the few-shot inputs in your LLM scoring prompts.
Our guides on automated QA for call centers and call center QA software selection cover implementation patterns in depth.
Defining key metrics for your QA scorecard
Monitoring agent script compliance
Script compliance tracking checks whether required phrases were spoken, in the correct order, by the correct speaker. For regulated industries, these are compliance requirements with direct legal exposure, not coaching opportunities. An automated scoring model can flag every call where the disclosure was absent, partial, or delivered by the customer rather than the agent.
That last scenario, the customer repeating back the disclosure, is exactly why speaker diarization is the prerequisite for compliance scoring. Without knowing who said what, a transcript-based compliance check produces false passes. Our speaker diarization capability, powered by pyannoteAI's Precision-2 model, provides that attribution in async workflows.
Evaluating agent empathy and tone
Soft-skill scoring is where automated QA draws the most skepticism, and the concern is legitimate. The key is precision about what automated scoring can and cannot do. Our transcription pipeline provides text-based sentiment inference, running NLP models against the transcript to score what was said. It does not provide acoustic emotion detection, which analyzes vocal characteristics like pitch, prosody, and speech rate to infer tone independent of word content. For most QA applications, text-based sentiment is sufficient to identify escalation risk and flag calls for human coaching review, and it avoids the privacy and compliance complexity of analyzing vocal biometrics.
Verifying transcription for scoring
A correct scorecard result requires knowing which speaker said each phrase. If diarization is incorrect and the transcript assigns the customer's escalation language to the agent, a compliant interaction scores as a failure. Speaker diarization is available in async workflows via our pre-recorded transcription API and is not available in real-time workflows. Contact centers running automated QA should plan their pipeline around post-call batch processing, with speaker attribution handled in post-processing for final scorecard accuracy.
Automating review with high precision
When you feed a transcript into an LLM for scoring, transcription errors compound in a non-linear way. A substitution of a common filler word has minimal impact on the LLM's interpretation, but a misheard account number, a dropped compliance phrase, or a garbled entity causes the LLM to either miss a violation or flag a false positive. At scale, even a modest error rate in the underlying transcript generates enough noise in the scoring output to undermine coaching credibility with agents.
Designing your call center QA scorecard
Build your scorecard in five steps:
- Define core metrics: Align your metric set to operational KPIs (FCR, CSAT, AHT). Industry guidance consistently favors compact, focused scorecards over long checklists. That's enough criteria to cover your key risk areas without diluting evaluator or LLM scoring consistency.
- Separate binary from scaled criteria: Binary criteria (disclosure read, prohibited language absent) can be scored deterministically by an LLM. Scaled criteria (empathy, tone appropriateness) require anchored examples at each score level in your rubric.
- Define critical failures: Specify which violations (PCI breach, HIPAA disclosure failure, prohibited language) produce automatic scorecard failure regardless of overall score.
- Write LLM-parseable rubrics: For each criterion, provide exact phrases or paraphrases that count as passing (binary) or one anchored example per scale point (scaled). Calibrate scoring consistency by building a multi-reviewer gold-standard sample, calls scored independently by at least two reviewers with inter-rater agreement measured, then compare automated model outputs against that baseline to identify and correct systematic drift before deploying at scale.
- Simplify for production: Remove redundant criteria, consolidate overlapping categories, and ensure each criterion is independently evaluable without requiring context from another metric.
The critical link between audio quality and insight
The hidden costs of manual QA
The industry benchmark for manual QA coverage is as little as 1% to 2% of total call volume, reflecting reviewer bandwidth constraints rather than operational design. Beyond cost, the deeper problem is the sample itself. Manual sampling, even when structured to include escalations and randomly selected calls, leaves entire agent cohorts invisible. A top performer never reviewed is never coached. A systemic compliance gap in one BPO region never sampled never triggers a remediation.
Organizations that move from sampling to full coverage consistently discover compliance gaps and coaching opportunities that were invisible under the sample. The cost of that blind spot does not appear as a line item. It shows up in FCR, CSAT, and cost-per-contact, spread across enough metrics that it rarely triggers a direct conversation until a compliance audit surfaces it.
Garbage in: why raw data accuracy matters
Our Solaria-3 model is specifically built for real-world contact center audio. It ranks first on real customer English audio at 9.6% WER on a human-annotated dataset of real customer calls, per our benchmark methodology, and #1 on Earnings22 financial calls at 6.4% WER, ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics. The Solaria-3 introduction video covers benchmark methodology and real-world audio performance across noisy, accented, and conversational conditions.
The table below shows how Solaria-3 compares against direct competitors on the benchmarks most relevant to contact center QA:
| Benchmark |
Solaria-3 |
Top competitor (per benchmark) |
Gap |
| Real customer English audio (WER) |
9.6% |
ElevenLabs 9.9% |
#1 |
| Earnings22 financial calls (WER) |
6.4% |
AssemblyAI 6.9% |
#1 on Earnings22 |
| Switchboard conversational (WER) |
33.9% |
ElevenLabs 55.2% |
#1, only model under 35% |
Why transcript precision drives QA metrics
Every downstream system, including your CRM sync, coaching scorecard, compliance flags, and sentiment analysis, operates on the transcript. A misheard entity name corrupts the CRM entry. A dropped disclosure phrase produces a false pass on your compliance check. A speaker misattribution scores the wrong agent on the wrong behavior.
This is why Aircall, which now handles over 1 million calls weekly through our infrastructure, made the switch. Every QA signal from AI summaries to sentiment scores to agent coaching runs on the transcript layer. When that layer is wrong, everything downstream is wrong, and the damage is invisible until it surfaces as a disputed QA score or a failed compliance audit.
How intelligent transcripts power objective scoring
How to achieve 100% call review volume
Batch (async) transcription is the foundation of economically viable 100% QA coverage. A POST request to our pre-recorded transcription endpoint processes approximately 60 seconds per hour of audio, returning structured JSON with speaker-attributed segments, timestamps, entities, and sentiment scores.
Our pricing model makes the arithmetic straightforward. At our Growth tier, async transcription starts at $0.20/hr based on audio duration, with diarization, sentiment analysis, translation, named entity recognition (NER), and summarization all included in the base rate. There are no separate line items for the features that make QA automation work. Deepgram's add-on model bills sentiment analysis, topic detection, and summarization as separate token-based charges (per Deepgram's public pricing), and AssemblyAI's features stack separately at effective rates approaching $0.45/hr for a full QA-grade feature set on their base-tier model. At 50,000 hours of annual call volume, the bundling difference compounds materially against your cost-per-contact target.
Handling multilingual QA at scale
BPO (Business Process Outsourcing) operations in Southeast Asia, South Asia, and Latin America introduce languages and dialects that most speech-to-text (STT) providers do not reliably support. When agents switch languages mid-conversation (code-switching from English to Tagalog, from English to Hindi, or from Spanish to English), single-language models either fail silently or return garbled output, breaking compliance scoring for the affected segments.
Our Solaria-1 model covers 100+ supported languages. True mid-conversation code-switching detection is built in, so when a bilingual agent shifts to a second language, the transcript stays continuous rather than breaking. For European business audio (EN, FR, DE, ES, IT), Solaria-3 is the right model. For global multilingual BPO, Solaria-1 handles the breadth.
Automated compliance and disclosure tracking
Named entity recognition identifies and classifies key information in transcripts, including names of people, organizations, and locations, with additional entity types such as product codes and account references configurable via custom vocabulary. NER runs through our audio intelligence pipeline when enabled in your API request and is included in the base rate with no add-on pricing.
PII (Personally Identifiable Information) redaction requires explicit configuration and is not enabled by default. Apply it selectively: redact PII in the output stored to your data warehouse, but preserve it in the transcript fed to your LLM scoring model so soft-skill criteria (like whether the agent used the customer's name) can be evaluated correctly. Our compliance and certification guide covers the configuration approach for regulated environments.
Automating QA with CRM and WFM loops
The integration pattern is straightforward: call recording triggers a POST to our pre-recorded API with callback: true and a callback_config object specifying your webhook URL, we POST the structured JSON to your webhook once processing completes (typically within 60 seconds per hour of audio), your backend runs the transcript against the LLM scoring prompt with your rubric, and scores write to your CRM or WFM coaching queue. Our integration recipes guide covers Salesforce, HubSpot, Airtable, and Slack destination patterns. If you're migrating from an existing provider, our real-time migration guides for Deepgram and AssemblyAI cover WebSocket endpoint mapping and authentication differences for streaming workflows. For the async pre-recorded surface, authentication follows the same API key pattern documented in our getting started guide. Most teams are live in under a day with direct Slack access to our engineers.
How transcription feeds your QA scorecard in production
Linking speech data to QA workflows
Once a call is transcribed, our API returns a structured JSON payload with speaker-attributed transcript segments, timestamps, and sentiment scores. Here is what that output looks like for a contact center QA workflow:
```json
{
"transcription": {
"utterances": [
{
"start": 0.15,
"end": 4.20,
"channel": 0,
"speaker": 0,
"text": "Thank you for calling support. My name is John. How can I help you today?"
},
{
"start": 4.85,
"end": 9.10,
"channel": 0,
"speaker": 1,
"text": "Hi John, I am calling because my billing statement has an incorrect charge of fifty dollars."
}
]
},
"sentiment_analysis": {
"results": [
{
"text": "Thank you for calling support. My name is John. How can I help you today?",
"sentiment": "neutral",
"start": 0.15,
"end": 4.20,
"channel": 0,
"speaker": 0
},
{
"text": "Hi John, I am calling because my billing statement has an incorrect charge of fifty dollars.",
"sentiment": "negative",
"start": 4.85,
"end": 9.10,
"channel": 0,
"speaker": 1
}
]
}
}
```
Your scoring model reads speaker: 0 utterances for agent compliance criteria and cross-references speaker: 1 entries in sentiment_analysis.results for customer sentiment trajectory. The structured output feeds into a standard LLM prompt, a webhook, or a direct API write to your CRM or WFM system without a separate enrichment hop.
Security protocols for call records
Our compliance stack covers SOC 2 Type II, ISO 27001, HIPAA, GDPR, PCI, and HDS (Hébergeur de Données de Santé), France's mandatory framework for hosting personal health data. For clinical call centers and healthtech platforms operating under European health data law, HDS certification removes the compliance blocker that previously ended the vendor conversation before pricing came up.
On our Growth and Enterprise plans, customer audio is never used for model training, and no opt-out action is required. On our Starter plan, customer data can be used for training by default. For operations handling regulated financial or health data, plan tier selection is a compliance decision as much as a pricing one. We do not support on-premises or air-gapped hosting on any plan. Deployment runs on dedicated cloud clusters in EU and US regions, with multi-region data residency configurable to your geographic footprint. Full compliance documentation is available through our compliance hub.
How Gravite cut review time by 93%
Gravite is a French CCaaS quality-monitoring platform processing 50,000 hours of audio per year. Before adopting automated transcription, their QA team spent approximately 15 minutes reviewing each call, which made 100% coverage operationally impossible and forced reliance on manual sampling.
After deploying our CCaaS audio infrastructure, they cut per-call review time from 15 minutes to 1 minute, a 93% reduction. Three requirements drove the selection: French-first accuracy on production contact center audio, European data residency, and Solaria-3's benchmark performance on telephony audio. They run zero fallback providers.
The operational shift was from broad sampling to exception-based review. Automated transcription and scoring covers all 50,000 hours annually. The QA team's function changed from reviewing a random or stratified sample to investigating flagged calls, the interactions where scores fall below threshold or where specific compliance criteria fail. Human review became targeted and high-value rather than broad and bandwidth-limited. The Gravite case study details the before/after workflow and how the team restructured their coaching cadence around automated flagging.
Start with €50 in free credits and have your integration in production in less than a day. The blind STT comparison tool strips out provider branding so you pick the better transcript before seeing who produced it, a useful gut check before running a reproducible benchmark on your full audio distribution.
FAQs
How many calls should a manual QA team review?
Manual QA teams typically review as little as 1% to 2% of total call volume due to reviewer bandwidth constraints, leaving the vast majority of interactions unreviewed. This limited coverage misses systemic compliance risks and coaching opportunities that only surface when reviewing all calls.
What is the difference between a QA scorecard and a QA rubric?
A QA scorecard defines the metrics and their weights, and it is the measurement output. A QA rubric defines the grading criteria for each metric, the evaluative framework that tells a reviewer or automated scoring model whether a behavior was present, absent, or partially met.
How many criteria should a QA scorecard track?
Industry guidance consistently favors compact, focused scorecards over long checklists. A narrower, well-defined metric set covers your key risk and coaching areas without diluting scoring consistency across evaluators and LLM prompts, and it is easier for agents to internalize and act on.
What transcription accuracy is needed for automated scoring?
There is no universal threshold, because the acceptable error rate depends on the scoring criteria. For binary compliance checks, even a single dropped phrase produces a false pass, so accuracy requirements are effectively zero tolerance on disclosure-critical segments. For soft-skill criteria, a higher error rate is tolerable if entity names and sentiment-bearing phrases are preserved. On real English customer calls, Solaria-3 ranks #1 ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics.
What is the difference between Solaria-1 and Solaria-3 for contact centers?
Solaria-3 ranks #1 on real-world European and English business audio, the most accurate on Earnings22 financial calls and Switchboard conversational audio, ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics. Solaria-1 covers 100+ languages and true mid-conversation code-switching for global BPO operations where language breadth matters most.
Does your transcription API support real-time speaker diarization?
No, speaker diarization is an asynchronous, post-call capability powered by pyannoteAI's Precision-2 model. For real-time workflows, speaker attribution must be handled in post-processing to maintain high scoring accuracy on the final QA record.
Where is customer audio data hosted and processed?
We deploy exclusively on dedicated cloud clusters in EU and US regions, and on-premises or air-gapped deployment is not available on any plan. EU-based operations can configure their regional cluster to meet data residency requirements.
Is customer audio used to train your speech models?
On our Growth and Enterprise plans, customer data is never used for model training, and no opt-out action is required. On our Starter plan, customer data can be used for training by default, making plan tier selection a compliance decision for regulated industries.
Key terms glossary
Word error rate (WER): The standard metric used to measure speech-to-text accuracy by calculating the percentage of insertions, deletions, and substitutions in a transcript compared to a reference transcription. The audio condition (conversational, financial, noisy) must always be stated alongside the figure.
Diarization error rate (DER): The metric used to evaluate speaker diarization performance, calculated by summing the duration of three error types: speaker confusion, false alarms, and missed detections, divided by the total reference length. High DER directly breaks compliance and soft-skill scoring in automated QA pipelines.
Speaker diarization: The process of partitioning an audio recording into segments associated with individual speakers, allowing a system to determine who spoke when. In contact center QA, diarization is the prerequisite for attributing compliance disclosures, sentiment changes, and talk-time to the correct agent or customer.
Code-switching: The practice of alternating between two or more languages or dialects within a single conversation. In multilingual BPO environments, code-switching causes single-language ASR models to produce gaps or errors in the transcript that break downstream scoring for those segments.
HDS certification: Hébergeur de Données de Santé is France's mandatory security framework for hosting personal health and patient data. HDS certification qualifies clinical call center and medical dictation deployments under European health data law, alongside SOC 2 Type II, HIPAA, and GDPR compliance.
First contact resolution (FCR): The percentage of customer issues resolved on the first interaction without requiring a callback or follow-up. FCR is the most direct operational output of QA coaching effectiveness and is used to benchmark agent and team performance across contact center operations.
Average handle time (AHT): The average duration of a customer interaction including talk time, hold time, and post-call work. AHT is a primary cost-per-contact driver and is monitored alongside FCR because improvements in one often affect the other.
Customer satisfaction score (CSAT): A metric measuring customer satisfaction with a specific interaction, typically collected through post-call surveys. CSAT is a key performance indicator for contact center quality and is often tracked alongside FCR and AHT to measure the impact of QA coaching.
Business process outsourcing (BPO): The practice of contracting specific business operations (such as customer service, technical support, or sales) to a third-party provider, often in a different geographic region. BPO contact centers frequently handle multilingual customer interactions and require speech recognition that supports code-switching.
Speech-to-text (STT): The technology that converts spoken language in audio recordings or live speech into written text transcripts. Also referred to as automatic speech recognition (ASR), STT is the foundational layer for all downstream contact center AI applications including QA scoring, sentiment analysis, and compliance monitoring.
