Contact centre speech analytics transcribes every call and analyses the text: why customers call, whether agents followed required scripts, where calls go wrong, and what each conversation should add to the CRM. It replaces manual quality review of a small sample with coverage of all calls. The results are only as good as the transcripts, and the words that drive analytics, product names, order numbers and key phrases, are the ones general models miss most.
Key takeaways
- The core shift is from sampling to coverage: every call transcribed, searchable and scored, instead of a handful per agent reviewed by hand.
- Most contact centre audio is 8 kHz telephony. Record in dual channel if you can; it solves speaker attribution for free.
- Post-call batch transcription covers analytics and QA. Real-time streaming is needed only for agent assist and live voice agents, and usually costs more.
- Redact card data and other personal data before transcripts reach analytics tools and language models.
- The EU AI Act prohibits emotion recognition in the workplace, which rules out inferring agents' emotions from their voices, with narrow exceptions.
What speech analytics does#
- Contact drivers. Classify calls by reason ("where is my order", "cancel subscription", "billing error") and track volumes over time.
- Quality assurance. Score calls against a QA form: greeting, identity check, resolution, closing. Automated scoring lets supervisors focus on flagged calls.
- Compliance. Check that required disclosures were read, such as a recording notice, a cooling-off period or a risk warning, and that prohibited phrases were not used.
- Summaries and wrap-up. Generate a call note for the CRM, replacing manual after-call work. See conversation summarisation.
- Agent assist. In real time, surface a knowledge article when a customer mentions a product, or remind the agent of a missing step.
- Search. Find every call in which a product defect or competitor was mentioned last week.
The pipeline#
Contact centre audio#
Almost all calls pass through the telephone network, which means narrowband audio at 8 kHz, often G.711 encoded, sometimes compressed further by VoIP codecs. See telephony transcription. Other features specific to contact centres:
- Dual-channel vs mono. Many recording platforms can store the customer and agent on separate channels. Transcribing each channel separately gives near-perfect speaker attribution. Mono recordings need diarisation, which fails most often on short turns such as "yes" and "okay".
- Non-speech segments. IVR prompts, hold music and transfer tones. Models must neither transcribe music as words nor hallucinate during long holds.
- Transfers and conferences. A call may have three or four parties in sequence, each needing a label.
- Accents and languages. Pan-European centres handle several languages, often on the same queue. See spoken language identification.
Why transcript accuracy drives analytics quality#
Analytics is built on words. If the transcript misses them, every downstream count is wrong in a way that no dashboard reveals.
- Product and plan names. the Unlimited Max bundle transcribed as the unlimited Macs bundle disappears from the product report.
- Order and case numbers. Linking a call to its order requires the exact identifier. See structured identifiers.
- Compliance phrases. A disclosure check that looks for "this call may be recorded" fails when the transcript reads "this call maybe recorded" unless matching is fuzzy, and fuzzy matching lets through variants that do not meet the script.
- Negation. "I'm not happy with the resolution" and "I'm happy with the resolution" feed opposite satisfaction scores.
Evaluate vendors on keyword accuracy for your own product list and on identifier accuracy, not only WER. Classic telephone benchmarks such as Switchboard are conversational but decades old and American English only.
Privacy, payments and AI Act limits#
Payment card data#
PCI DSS prohibits storing sensitive authentication data, such as the card verification code, after authorisation, and requires protection of stored card numbers. Contact centres typically use pause and resume, DTMF masking so customers enter digits on the keypad, or automated redaction of audio and transcript.
GDPR#
Customers must be informed that calls are recorded and why (GDPR Articles 13 and 14). Each purpose, such as quality monitoring, training, analytics or feeding a language model, needs a lawful basis, and legitimate interests requires a balancing test. Agents are data subjects too. In several Member States, monitoring employees requires consultation or agreement with employee representatives; in the Netherlands, for example, the works council has a right of consent over systems for monitoring staff under the Works Councils Act (WOR, Article 27).
EU AI Act#
Article 5(1)(f) of Regulation (EU) 2024/1689 prohibits AI systems that infer the emotions of people in the workplace, except for medical or safety reasons. This has applied since 2 February 2025 and covers analysing agents' voices for stress, frustration or enthusiasm. Emotion recognition applied to customers is not prohibited but is listed as high-risk in Annex III and carries transparency duties. Analysing what was said, such as topics, disclosures and resolution, is not emotion recognition. See the EU AI Act and speech recognition.
How Reson8 approaches this#
Customer support is one of the industries Reson8 serves. Batch transcription costs 1 credit per minute and real-time streaming 2, so post-call analytics can run at half the cost of live transcription. Patterns takes an expected identifier format and biases the model toward it; on order numbers, WER fell from 8.0% to 1.5%. Product catalogues can be supplied as text through domain adaptation, taking effect in under 60 seconds. Audio is processed on EU GPUs with zero retention.
Common questions#
Do we need real-time transcription?
Only for agent assist or voice agents. QA, compliance, topic analysis and summaries all work post-call in batch, which is cheaper and can use the whole call as context.
What share of calls should be transcribed?
For analytics, all of them; partial coverage reintroduces the sampling problem. Exclude calls where you lack a lawful basis or where customers objected.
Can we use transcripts to train agents?
Yes, if the purpose is covered by your privacy notice and lawful basis. Redact personal data first, and involve employee representatives where national law requires.
Is it legal to score agents automatically?
Automated QA on what was said is common, but decisions with significant effects on an employee based solely on automated processing are restricted by GDPR Article 22. Keep a human in the loop for performance decisions.
References#
- Godfrey, J. J., Holliman, E. C., & McDaniel, J. (1992). SWITCHBOARD: telephone speech corpus for research and development. ICASSP.
- ITU-T Recommendation G.711: Pulse code modulation (PCM) of voice frequencies. International Telecommunication Union.
- PCI Security Standards Council. Payment Card Industry Data Security Standard (PCI DSS) v4.0. pcisecuritystandards.org.
- Regulation (EU) 2016/679 (General Data Protection Regulation). EUR-Lex.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5 and Annex III. EUR-Lex.