Quality
Speech analytics for QA: every call instead of 1-2%
Manual call monitoring covers 1-2% of calls, and problems slip through. Speech analytics checks every conversation against your scorecard, and supervisors review only the calls it flags.
A QA specialist typically listens to a few calls per agent per month and fills out a scorecard. An agent handles about 50 calls a day – over 1,000 a month. Listen to 10-20 of them, and you've checked less than 2%. The rest slips by, and agents see scoring as a lottery. Speech analytics transcribes and checks every call.
Why random sampling isn't enough
- Coverage: a couple of calls won't tell you how an agent works all month.
- Bias: reviewers pick long, unusual, or simply easy-to-find calls.
- Delay: a mistake is found two weeks later, after it has been repeated hundreds of times.
- No big picture: a random sample won't show why customers call or what they complain about.
- Disputes: the agent says the reviewed call wasn't typical – and is often right.
What it costs to check everything by hand
Take a team of 20 agents: 50 calls a day, 22 workdays – 22,000 calls a month. A QA specialist spends about 8 minutes per call: 4 minutes to listen and the same to fill out the scorecard. Checking 2% means 440 calls and 59 hours – a third of one person's working month. Checking everything means 2,933 hours, the work of 17 full-time people. Speech analytics does this work without new hires.
What speech analytics does
- Transcribes the call and separates what the agent and the customer said (this requires dual-channel recording, with the agent and the customer on separate tracks).
- Finds words and phrases: the greeting, required disclosures, banned expressions, competitor names.
- Detects silences, interruptions, crosstalk, and long holds.
- Sorts calls by topic and outcome.
- Flags customer frustration and signs of conflict.
- Scores each call against your scorecard and flags calls worth a listen.
Chats and email are even easier to check: the text is already there, no transcription needed. The same rules work across all channels.
How it works under the hood. Speech recognition turns the recording into text and timestamps every word – that's how silences and interruptions show up. Then the text is checked two ways: rules look for specific phrases and their variations, and a language model judges meaning – for example, whether the agent uncovered the customer's need even if they put it in their own words. That points to the main source of errors: with poor audio, recognition misses words, and the system decides the phrase was never said.
How to turn your scorecard into rules
- List the required steps of a call: greeting, customer verification, needs discovery, offer, summary, and closing.
- For each step, write down sample phrases and their variations: "How can I help you?", "I'm listening", "Tell me what happened."
- Assign weights: skipped verification is critical, a missing summary at the end is a lower-priority fix.
- Run scoring on all calls and see what share of conversations skips each step.
- Listen manually only to flagged calls and calls with low scores.
Because of recognition errors, automated scoring gets individual calls wrong – more often with background noise and vaguely written rules. Use it to find recurring mistakes and pick calls for review. A person makes the call on disputed cases.
Required disclosures and legal requirements
In banking, insurance, telecom, and healthcare, some phrases are required and others are prohibited. A missed disclosure can mean a fine or a lost dispute with a customer.
- Check that the agent gave the call recording notice – several US states require consent from all parties – and, where needed, got the customer's consent to process personal data.
- Flag prohibited promises like "you're definitely approved" and rudeness.
- Make sure the agent verifies the customer before discussing an account, contract, or diagnosis.
- Keep searchable transcripts – you need them for audits and complaint reviews.
The score comes in shortly after the call. With random sampling, an agent taking 50 calls a day can have 500 conversations in the two weeks before a review – for example, all without a required disclosure. Here, the supervisor finds out the same day and fixes it.
Why customers call: topics and trends
Speech analytics also answers a manager's biggest question – why customers call.
- The share of each topic by day, week, and campaign.
- Sudden spikes in a topic: website checkout broke, a delivery is late, customers don't understand a new bill.
- Repeat calls and their causes.
- Topics you can hand off to a voice AI agent or the IVR.
- Insights for product, marketing, and logistics teams.
Topics show where the money goes. Say "where's my order" makes up 15% of calls – 3,300 of 22,000 a month at 4 minutes each, or 220 hours of talk time. Agents spend only about 60% of their time talking, so that one topic takes about 370 paid agent hours a month. You could send order status by text message or hand the question to an AI agent. The AI agent ROI calculator shows how much that would save.
When a complaint comes in, you find the right call by keyword search in minutes, not by asking agents to remember.
From scores to coaching
- Each agent gets a dashboard with trends, not one random score.
- Review a specific call: the transcript in front of you, the missed step highlighted.
- Top agents' best calls become training examples for new hires.
- A few weeks after coaching, check whether the behavior changed.
- The AI assistant for agents reminds agents about a missed step during the call itself.
What changes for supervisors and agents
The QA team stops scoring random calls and works on recurring mistakes across all conversations. QA specialists stay: they coach agents and review flagged calls instead of hunting for something to listen to.
- Supervisors start the day with a list of calls that need attention.
- Agents see their own trends and improve before a formal review.
- Fewer disputes about scores: every call is checked by the same rules.
- Managers see quality next to service level, handle time, and repeat calls.
From day one, explain to agents what the system checks, how the score is calculated, and where to see their results. When people see that the goal is to help, not punish, they check their dashboards without being asked.
Where to start
- Work out what share of calls you check now: calls reviewed in a month divided by all calls.
- Keep 5-8 critical items from your current scorecard and start with one queue.
- Compare automated and manual scores on a sample and adjust the rules where they disagree.
- Add a vocabulary list: your product names, brands, and industry terms – recognition makes fewer mistakes that way.
- Agree with the team that results are mainly for coaching.
- After a month, find the most expensive call topics, as in the example above, and add the other queues, chat, and email.
In CONTACT.WORK, speech analytics works with call recordings from the same platform: scores and topics show up in the same reports as service level and handle time. Recordings and transcripts are stored in our cloud or on your servers. The platform is free; speech recognition and the language model are paid separately – or they run on your own GPU or Mac mini, and then recordings never leave your infrastructure.