Contact-center speech: quality and compliance on every call
Every call is recorded, quality hears two in a hundred, and off-the-shelf speech-to-text mangles your product names.
- Kinds of AI
- SLMsAgentic AI
- Who decides
- Team lead
- Where it runs
- Your servers or your cloud account
SLM · quality score
SLM · scorecard, with the moment
- metGreeting and identity check00:12
- metKVKK notice read00:31
- not metFee refund offered without an eligibility check03:58policy
- Compliance exceptions1to the compliance analyst queue
- SLM
- Agent
Written limitThe score is advisory. No speaker identification.
DecidesTeam lead
1. The desk
Today
- TranscribeGeneric Turkish transcription gets product names, campaign codes, and IBANs wrong.
- ScoreQuality listens to a small random sample of calls.
- CheckCompliance items (the KVKK notice read, no rate promises, the complaint logged) are checked by sampling.
- CoachThe team lead's coaching arrives weeks late.
With the system
- TranscribeThe speech model, fine-tuned on your call audio and terms, transcribes every call with timestamps and masks personal data in text and audio.SLM
- ScoreThe SLM scores every call against your scorecard and compliance rules, quoting the words and the moment behind each finding.SLM
- CheckThe agent puts each compliance exception in the compliance analyst's queue and adds the call to the coaching file of the call-center agent.Agentic AI
- CoachThe team lead sees the coaching point in context and decides every score that affects a person's review, as before; the model's score is advisory.Person decides
What lands on the desk
Every call arrives scored, with the moment of each finding and the words that support it. The team lead sees the coaching point in context, and compliance sees only the calls with an exception.
Speech model · transcript, redacted
…kart numaranızın son dört hanesi [REDACTED]…
SLM · quality score
SLM · scorecard, with the moment
- metGreeting and identity check00:12
- metKVKK notice read00:31
- not metFee refund offered without an eligibility check03:58policy
- Compliance exceptions1to the compliance analyst queue
SLM · coaching note draft, for the team lead
Check eligibility before promising a refund.
- SLM
- Agent
Written limitThe score is advisory. No speaker identification.
DecidesTeam lead
2. The system
For your technical team — skip it if you like.
How it works
In
- Call audio
- One channel per speaker where possible
Speech model · transcribes
Speech recognition (Whisper)
Whisper, LoRA-tuned on your Turkish call audio and terms, 8 kHz
Reads
- Everything under In
Out
- A transcript with timestamps
- ID, IBAN, card, phone, address redacted
- Redacted in text and audio
SLM · understands
Quality and compliance reader
Post-trained on your quality scorecard and compliance rules
Reads
- The transcript
- The quality scorecard
- The compliance rules
Out
- A score per call
- Quotes and timestamps as evidence
- Compliance exceptions
Agent · prepares
Coaching and exception agent
Fixed workflow
Reads
- The scores and exceptions
Out
- A coaching file per agent, with examples
- A compliance exception queue
Decision
Team lead and compliance analyst
Reads
- Any score that affects a person's review
- Every exception
Out
- The team lead makes the review decision
- The compliance analyst confirms each exception
How the agent's work is arrangedFixed workflow
Fixed workflow
Every scored call goes through the same steps (queue the exceptions, add to the coaching file), so no routing or loop is needed.
How the models are trained and tested
Your expert writes the test first. A step the test does not need is skipped, and we say why.
Show the training steps
Speech recognition model (ASR)
Test set
Held-out calls from different days and call-center agents; word error rate (WER) and an error rate on your own term list.
Dataset
20–100 hours of your call audio, transcribed by people to a style guide (numbers, brand terms), cut to 30 seconds or less, redacted before it leaves the secure zone, and spread across products, channels, and accents.
Continued pre-training (CPT)Skipped
Skipped: the base model already knows Turkish speech; your terms are taught by fine-tuning.
Supervised fine-tuning (SFT)
LoRA fine-tuning, plus text-normalization rules for numbers and codes.
Reinforcement learning (RL)Skipped
Skipped: fine-tuning on transcripts made by people covers the task, and no further reward is needed.
Licenses
Whisper is MIT-licensed; speaker separation (diarization with pyannote, which splits who spoke when, never who the speaker is) only if channels are mixed.
Quality and compliance SLM
Test set
Calls scored by your senior evaluators; the model's agreement with them is measured.
Supervised fine-tuning (SFT)
On your scorecard and compliance rules.
Reinforcement learning (RL)Skipped
Skipped: the scorecard is taught by fine-tuning.
Where the agentic system stops
The team lead decides every score that affects a person's review; the model's score is advisory.
Show its tools, limits, and record
- Tools and access
- Read transcripts and scores; write coaching drafts and exception tickets.
- May not
- Change a call-center agent's rating, take disciplinary action against staff, or contact a customer.
- Who decides and the written limit
- Team leadThe team lead decides every score that affects a person's review; the model's score is advisory. A compliance analyst confirms every exception.
- Stops when
- Transcription confidence is low on a scored item: a person listens to the call.
- Record
- Model versions, evidence quotes, and the team lead's decisions.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Quality coverage from a sample to every call.
- Faster compliance findings.
- Coaching aimed at the moments that matter.
What your institution measures
- Quality coverage
- Word error rate and error rate on your terms
- Agreement between the model and your evaluators
- Compliance exceptions per 1,000 calls
- First-contact resolution
- Average handling time
Fine-tuned on a small Turkish speech dataset (Mozilla Common Voice), Whisper-Small reached a word error rate of 0.16 on that dataset, against 0.28 for Wav2Vec2-XLS-R-300M; on call-center recordings outside its training data, both models' error rates were high.
Measured in the Decision Blueprint
Today's quality coverage and the current transcription error on a sample of your calls set the baseline; your institution sets the target.
Conditions
The rules, the data, and the known error, with its guard.
- No speaker identification and no voiceprints: KVKK treats voice timbre as biometric, special-category data.
- Recording needs a lawful basis and notice; continuing a call is not by itself explicit consent.
- Training on recordings is a new purpose that your legal owners decide.
- Scoring staff needs transparency to employees. Audio stays in your environment.
Start here
We take a sample of your calls and your scorecard, measure today's transcription error on your terms, and check how your senior evaluators agree before any model is trained.
Similar desks
- SLMsAgentic AI
TodayNew call-center agents take months to be productive and put customers on hold to search the knowledge base.
With the systemAnswers drafted while the customer is on the line.
Call-center agent decides
- SLMsAgentic AI
TodayOne team answers written complaints for several clients, each with its own rules, tone, and deadlines.
With the systemEach reply drafted in the client's own rules and tone.
Team lead decides
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision is enough; no confidential records.
https://eqlx.ai/use-cases/contact-center-speech · Printed copy; the website is the current version.