Transcription Accuracy and What Degrades It

Every claim a speech analytics platform makes rests on the transcript. Accuracy on your calls is a different number from the one in the brochure, and measuring it is the first thing to do.

How accuracy is quoted

Word error rate is the standard measure: the proportion of words substituted, deleted or inserted relative to a human reference transcript.

Vendor figures typically come from clean read speech — benchmark corpora with good microphones, standard accents and no crosstalk.

Contact centre audio is the opposite of that: compressed telephony, background noise, overlapping speakers, varied accents, domain vocabulary, and callers on mobile phones in cars.

The gap between benchmark and production is frequently substantial, and it is not disclosed because it depends entirely on your audio.

What degrades it, in rough order

Audio quality. Codec compression, packet loss, poor headsets. This is the largest single factor and it is fixable.

Channel mixing. Both speakers on one channel makes diarisation — deciding who spoke — much harder and errors cascade. Stereo capture with agent and customer separated is the single biggest improvement available.

Crosstalk. People interrupting each other. Common in difficult calls, which are the ones you most want to analyse.

Accent and dialect, which has its own note because the disparity is systematic rather than random.

Domain vocabulary. Product names, drug names, part numbers, insurance terms. The recogniser has not seen them and produces something phonetically similar.

Proper nouns. Names of people, places and companies. Consistently the worst category.

Numbers and alphanumerics. Account references, postcodes, dates. High error rates and high consequence, because these are frequently what the call was about.

Background noise. Call centre floor noise, customer environments.

Measuring it on your own audio

Half a day of work and it changes how the whole deployment is evaluated.

Take thirty calls, stratified across accents, queues, and audio sources — landline, mobile, VoIP.

Get a human transcript of each, done carefully.

Compute word error rate against the machine transcript.

Then compute it separately by accent group, by queue, by audio source. The variation between groups is more informative than the average.

Also measure the terms that matter: how often are your product names, your key phrases, your compliance wording recognised correctly? A system with a good overall rate that fails on your twenty critical terms is unusable for compliance.

Improving it

Fix the audio first. Better headsets, stereo capture, higher-quality codecs where the telephony allows. This costs less than most software changes and helps everything downstream.

Custom vocabulary. Every serious platform accepts a list of domain terms, product names and acronyms. Populating it properly is high-return and frequently skipped.

Acoustic or language model adaptation where the vendor offers it, trained on your own audio.

Speaker separation at capture rather than by algorithm afterwards.

Then re-measure. An improvement claimed by a vendor should be visible in your own word error rate on your own sample.

What the accuracy figure implies

Above roughly 90 percent word accuracy: most analysis works, with care around proper nouns and numbers.

Around 80 percent: keyword detection is usable, categorisation is noisy, anything depending on exact phrasing is unreliable.

Below that: the transcript supports detection of common terms and little else, and any reporting built on it is measuring the recogniser as much as the calls.

Know which band you are in per accent group and per queue, because an average conceals a system that works well for some of your customers and poorly for others.

The honest position for a deployment

State the measured accuracy alongside every analytics output that depends on it.

A category count derived from a transcript with a 15 percent error rate is a count with a systematic bias toward whatever the recogniser hears well. Reporting it without that context is reporting the recogniser's preferences as a fact about your customers.

The vocabulary list

The cheapest accuracy improvement available and the one most often left empty.

Product and service names, including how customers mispronounce them.

Company and brand names, yours and competitors'.

Acronyms and internal shorthand, as they are spoken rather than written.

Regulatory and compliance phrasing, exactly as it must be said, because these are what compliance checks depend on.

Technical terms, part numbers and model codes.

Common customer phrasings for your products, which differ from your own terminology.

Build it from the calls, not from a product catalogue. Take a hundred transcripts, find the terms that came out wrong, and add them. Then re-measure and repeat.

Maintain it. New products, new campaigns and new terminology arrive continuously, and a vocabulary list from launch is stale within a year.