What word error rate actually measures
Every voice AI system, whether it is answering calls, transcribing them for a human agent, or feeding a summarisation model, starts with automatic speech recognition (ASR). Word error rate (WER) is the standard way to score how well that recognition works. It counts the substitutions, deletions and insertions a system makes against a verified human transcript, divided by the total number of words in that transcript. A 10% WER means the system got roughly one word wrong for every ten it processed.
WER is not a proxy metric bolted on for marketing. It is the oldest and most direct way to check whether a system heard what was actually said, and it predates the current wave of voice AI vendors by decades. The classic benchmark corpora for phone-quality audio, Switchboard and CALLHOME, have been used to score ASR systems since the 1990s, which is why researchers still reference them when building new evaluation methods (arXiv, 2024).
Why the number moves depending on who is speaking
WER is not a fixed property of a model. It moves with the audio. A 2024 study in JASA Express Letters tested OpenAI's Whisper model against a range of native and non-native English accents and speaker traits. The researchers found that native English accents were recognised more accurately than non-native ones, and that Whisper performed noticeably better on read speech than on natural conversational speech (Graham and Roll, JASA Express Letters, 2024). A contact centre call is conversational speech, by definition, and a large share of callers to any national support line will not speak in the accent the model was trained hardest on.
Background noise moves the number further. An academic benchmark of cloud speech-to-text services on French-language audio tested four major providers across signal-to-noise ratios from a clean 40 dB down to a noisy 0 dB. Microsoft Azure produced the lowest error rate on clean speech, at 9.09%, and held up relatively well as noise increased, but every provider's error rate rose as the audio got noisier (arXiv, 2021). None of that testing happened on a stage. It happened against recorded audio built to resemble what a phone call actually sounds like.
Where vendor claims and production numbers diverge
This is where the gap between the sales demo and the support queue opens up. A demo call is usually placed from a quiet office on a good headset, in the accent the salesperson happens to speak. A support queue is not that. Deepgram, an ASR vendor, publishes its own comparison putting its word error rate at 7.6% against 13.1% for Google Cloud's competing model on a defined test set (Deepgram, 2025, vendor benchmark). Numbers like this are worth noting, but they come from the vendor being compared, so they describe a controlled test condition, not a guarantee for any specific contact centre's callers.
AssemblyAI, another ASR vendor, states that leading APIs now reliably score below 5% WER on clean conversational English, while noting that real-world results depend on the audio conditions a system is actually deployed against (AssemblyAI, 2026, vendor study). Voicegain, a speech-to-text provider that focuses specifically on telephony audio, runs an annual benchmark built from a curated set of 8 kHz call centre recordings rather than broadband studio audio, comparing its own models against major cloud providers and OpenAI's Whisper on that lower-fidelity, compressed audio format that phone networks actually use (Voicegain, 2025, vendor benchmark). That distinction matters because 8 kHz telephony audio is a narrower, noisier signal than the broadband audio most public WER leaderboards are built on.
What to test before buying
The direct implication for a buyer is that a vendor's published WER, whatever it is, describes a specific test set under specific conditions. It does not describe your callers. Before signing, ask the vendor to run their system against a sample of your own recorded calls, not a reference dataset, and ask for the error rate broken out by call type, accent mix and background noise level rather than a single blended average. A single number hides exactly the variation described above.
It is also worth asking whether the quoted WER was measured on 8 kHz telephony audio or on higher-fidelity broadband recordings, since the two are not directly comparable. A model that scores well on a broadband benchmark can perform differently once run through the codec compression of a standard phone line.
What commonly goes wrong
The recurring failure is treating a single published accuracy number as a purchase decision rather than a starting question. A model can score well against a public benchmark and still misfire in a specific contact centre, because that centre's caller base skews toward accents, background environments or vocabulary the benchmark never tested. Word error rate is a real and useful measurement. It is just a measurement of the test it was run against, not of the phone line it will end up running on.
Sources: Graham & Roll, JASA Express Letters · Cloud speech-to-text benchmark, arXiv · ASR word error rate estimation, arXiv · Deepgram vendor benchmark · Voicegain 2025 benchmark · AssemblyAI vendor study