ASR Accuracy for Contact Centers: Why Demo WER Numbers Lie

Every ASR vendor publishes a word-error-rate (WER) number, and every one of those numbers is measured on conditions that don't resemble a contact center call. Clean wideband audio, a single accent, a quiet room, read speech rather than spontaneous conversation. Production contact center audio fails every one of those assumptions simultaneously, which is why a vendor's published 5% WER routinely becomes 15-20%+ on your actual call traffic.

What actually degrades accuracy, in order of impact

Telephony bandwidth. Phone audio is 8kHz (narrowband), roughly half the sample rate of the 16kHz+ audio most ASR benchmarks are measured on. This isn't a minor quality reduction — it removes frequency information the model may be implicitly relying on, especially for consonant sounds that distinguish similar-sounding words.

Background noise. Contact center audio comes from call center floors, cars, retail counters, and homes with TVs on. This is categorically different from the studio-quiet conditions of most benchmark corpora, and noise robustness varies enormously between ASR providers in ways that don't show up in a single blended WER number.

Accents and code-switching. Any ASR model's accuracy is a function of what accents were represented in its training data. If your caller base skews toward accents underrepresented in the vendor's training set, you will see materially worse accuracy than their published number — and if callers code-switch between languages mid-sentence (common in multilingual markets), most models degrade sharply, since they're implicitly assuming a single language per utterance.

Domain vocabulary. Product names, internal acronyms, and industry-specific terms are, by definition, rare in general-purpose training data. A caller saying your product name or a policy number will get transcribed as the nearest common-word approximation unless you've configured vocabulary boosting or custom terms — and this failure is invisible in generic benchmarks that don't use your vocabulary.

Cross-talk and overlapping speech. Real conversations include interruptions and simultaneous speech. Benchmark corpora are typically single-speaker or cleanly turn-taking. Overlapping audio is one of the largest sources of transcription errors in real contact center audio and one of the least represented in vendor marketing.

How to actually evaluate a vendor

Don't evaluate on their number. Build a real test corpus:

1. Pull actual call recordings from your existing IVR or a shadow-mode deployment — not synthetic test calls, not read scripts.
2. Segment by condition, not just an overall average: by accent group, by background-noise level, by call topic (some topics have heavier domain vocabulary). A single blended WER hides exactly the segments most likely to fail.
3. Test with your actual vocabulary — product names, account terminology, anything specific to your business — both with and without vendor-side custom vocabulary/boosting features enabled, since the uplift from that feature varies a lot by provider.
4. Include overlapping and interrupted speech in the test set deliberately, since it's undertested and commonly under-scored in generic benchmarks.
5. Re-test periodically, not just at vendor selection. Caller demographics shift, you launch new products with new vocabulary, and vendor models get silently updated — WER drift is ongoing, not a one-time gate.

The practitioner tip that actually matters here

Don't chase a single "best" ASR vendor. Accuracy leadership varies by condition — one provider might lead on noisy audio, another on accent diversity, another on domain-vocabulary handling with boosting enabled. If your call volume and cost structure support it, a router that picks ASR provider by detected condition (or that runs a cheaper model with a fallback to a more expensive one on low-confidence segments) will outperform any single "best" vendor choice on your actual mixed traffic. This adds real engineering complexity, so it's worth doing only once you've measured that the accuracy gap between providers is large enough to matter for your use case — see our voicebot cost comparison for how ASR choice feeds into the broader cost model, and our note on 8kHz telephony reality for how this connects to the migration rollout itself.

If you're evaluating ASR vendors and want help building a real test corpus instead of trusting a benchmark PDF, reach out.