Beyond Word Error Rate: Understanding the Shift from Traditional ASR to Neural Transcription in Contact Centers

A reference for contact center operators and CX platform users transitioning from legacy speech analytics to modern AI-powered transcription.

Introduction

When contact center teams move from legacy speech analytics platforms built on traditional automatic speech recognition (ASR) engines to modern AI-powered transcription — such as end-to-end neural models used in cloud CCaaS platforms today — one of the first questions raised is: “What is the word error rate?”

It is a reasonable question. WER has been the standard measure of transcription quality for decades. But on its own, it is not sufficient to evaluate a neural transcription system, and treating it as the sole or primary measure will lead you to incomplete conclusions about performance, set unachievable improvement targets, and miss what actually matters: whether the system correctly understood the call.

This article explains why. It covers how traditional ASR engines and modern neural transcription models are structurally different, why that difference makes WER alone an insufficient metric for evaluating neural systems, why traditional ASR is becoming a legacy paradigm, and what a more complete quality framework looks like in practice.

A note on sources: this article is based on best-effort research into publicly available literature, vendor documentation, and third-party benchmarks. Figures cited (WER percentages, training data volumes, and similar) reflect what has been published elsewhere, not measurements taken firsthand, and the sources are listed in the references at the end. One caveat applies throughout: published WER figures are typically produced under lab or benchmark conditions. Live performance on a specific contact center’s audio, call types, and network conditions will differ, sometimes considerably, and should not be assumed to transfer directly from the published number.

Part One: Two Fundamentally Different Architectures

How Traditional ASR Works

Legacy speech analytics platforms — built on engines such as Nuance, IBM Watson Speech, and proprietary enterprise ASR systems — use a classical automatic speech recognition architecture. This architecture has three distinct, separable components:

The acoustic model converts raw audio waveforms into phonemes: the smallest units of sound. It is trained on large volumes of labelled audio to recognise how sounds map to language.

The pronunciation dictionary is a lookup table that maps sequences of phonemes to words. If the acoustic model hears a particular sound pattern, the dictionary tells it which word that corresponds to.

The language model sits on top and uses statistical probability to determine which sequence of words is most likely given the preceding context. It gives the system its ability to resolve ambiguity: if the phonemes could spell either “their” or “there”, the language model picks the more probable option in context.

The crucial feature of this architecture is that each component is modular and separately tunable. A contact center operator can submit a list of domain-specific terms — product names, regulatory phrases, competitor names — and the system incorporates them into the pronunciation dictionary and language model. When a specific term is consistently misrecognised, a tuning cycle can correct it. The improvement is measurable, predictable, and customer-controlled.

This is why WER was a meaningful metric for legacy ASR: it was stable between tuning cycles, errors were systematic and reproducible, and there was a direct lever to pull to reduce it.

How Modern End-to-End Neural Transcription Works

Modern neural transcription engines — including OpenAI Whisper, Deepgram Nova, AssemblyAI Universal, Google Chirp, and Microsoft Azure Speech — are architecturally transformer-based end-to-end neural networks. They do not separate the acoustic model, pronunciation dictionary, and language model into distinct components. Instead, all three functions are baked into a single neural architecture trained simultaneously on hundreds of thousands of hours of multilingual audio.

OpenAI’s Whisper, for example, was trained on approximately 680,000 hours of audio sourced from across the internet. Deepgram’s Nova series was trained specifically on contact center and telephony audio at scale. Google’s Chirp model was trained on millions of hours across hundreds of languages. The scale and diversity of training data is precisely what gives these models their robustness — and precisely what makes them non-tunable in the traditional sense.

This has several important consequences:

There is no pronunciation dictionary to update. You cannot submit a vocabulary list and expect the model to treat those terms differently. The model generalises from patterns learned during training.

The language model is not separable. You cannot retrain or adjust the statistical language layer independently. The entire model is one artefact.

Errors are contextual, not systematic. Where a traditional ASR system will consistently misrecognise a specific term the same way every time (because the pronunciation dictionary is wrong), neural models produce probabilistic errors. The same term may be transcribed correctly in one call and incorrectly in another, depending on speaker accent, audio quality, surrounding context, and the model’s confidence at that moment.

The model improves via version releases, not customer tuning. When a provider releases a new model version, performance improves globally. Individual customers cannot accelerate or direct that improvement.

In short: traditional ASR and modern neural transcription are not competing implementations of the same thing. They are different paradigms of speech-to-text technology. Comparing their WER figures directly is not an apples-to-apples exercise. It is closer to comparing a mechanical watch to a quartz one by asking how often the escapement wheel turns.

Part Two: The Industry Transition — Traditional ASR Is Becoming a Legacy Technology

This architectural shift is not a matter of preference. It is an industry-wide transition already well underway, with significant consequences for how contact centers should think about quality management.

The Migration Away from Modular ASR

Nuance Communications — the most widely deployed enterprise ASR provider for nearly two decades — was acquired by Microsoft in 2022. Its on-premise ASR products (Recognizer and Vocalizer) reached end-of-life in 2025–2026, with Microsoft directing customers to migrate to Azure AI Speech. NICE CXone, which has historically used the Nuance ASR engine for IVR and speech analytics, is now transitioning its stack towards neural models through its Enlighten AI platform. Genesys and Avaya, which both licenced traditional ASR engines for on-premise deployments, now route new cloud deployments through Google CCAI or Microsoft Azure Speech — both of which are themselves neural architectures.

The pattern is consistent: the largest enterprise contact center platforms are retiring traditional modular ASR and replacing it with end-to-end neural models, either by building their own or by embedding cloud provider APIs.

Why the Transition Is Happening

The reasons are structural, not just performance-driven:

Neural models handle linguistic diversity better. Traditional ASR required a separate language pack, pronunciation dictionary, and acoustic model to be built and maintained for each language and dialect, each one its own standalone engineering effort. A neural model trained across many languages simultaneously can transfer patterns learned from higher-resource languages to lower-resource ones, and because it builds words up from smaller sound and sub-word units rather than matching against a fixed list, it can piece together unfamiliar terms, hesitations, and informal phrasing on the fly, rather than needing a separate configuration step for every variant.

Maintenance costs are lower. A modular ASR system requires ongoing tuning cycles, grammar file updates, and language model refreshes. A neural model is updated by the provider and deployed as a new version. For contact centers, this eliminates a class of operational overhead.

The downstream use case has changed. Traditional ASR was built to produce verbatim transcripts that human reviewers would read. Modern contact center AI is built to feed Large Language Model (LLM) pipelines that produce summaries, classifications, and sentiment scores. The neural architecture is better suited to this downstream use because it encodes contextual meaning rather than just surface-level phonemes.

Neural models are converging on a common accuracy floor. On contact center telephony audio at 8 kHz, leading neural models now achieve 6–14% WER depending on conditions. Traditional ASR platforms have historically reported 10–20% WER on the same audio type. The accuracy advantage of tunable traditional ASR has narrowed to the point where the tuning overhead is difficult to justify for most deployments.

What This Means in Practice

Contact center teams who move to neural transcription — whether through a CCaaS platform migration, a new cloud contact center deployment, or an API-based contact center speech analytics integration — are entering a fundamentally different operating model. The tuning levers they had with traditional ASR no longer exist in the same form. WER targets that were achievable under a modular system are no longer the right framework. The quality conversation needs to shift accordingly.

Part Three: The WER Problem

What WER Actually Measures

Word Error Rate is calculated as:

WER = (Substitutions + Deletions + Insertions) ÷ Total Reference Words

A substitution is a wrong word. A deletion is a missing word. An insertion is a word that appears in the transcript but was never spoken. The lower the WER, the closer the transcript is to a verbatim record of what was said.

This is a useful metric when verbatim accuracy is the goal: legal transcription, compliance recording, medical dictation. In those contexts, every word matters and a missed word can have consequences.

Why WER Alone Is Not Enough to Evaluate Neural Contact Center AI

Contact center speech analytics is not in the business of producing verbatim records. It is in the business of extracting insight: what did the customer call about, how was the issue resolved, what was the sentiment, were compliance scripts followed, what are the emerging call drivers this week?

None of these outcomes depend solely on every word being transcribed correctly. WER can tell you something useful about the raw input, but it says nothing on its own about whether the meaning and intent of the call were correctly understood, which is what the business actually needs from the system.

Consider a call about a lapsed direct debit. The agent explains: “Right, so what’s happened here is your standing instruction lapsed when the card on file expired last month.” The neural model mishears “Right” as “Alright” at the start of the sentence: a single-word substitution, a WER hit. It has no bearing on anything that matters. The post-call summary correctly records: Standing instruction lapsed due to expired card on file. Instruction reinstated for the customer. The system understood the call correctly. Judged on WER alone, the transcript looks flawed; judged on outcome, it is a pass.

Now consider the reverse. A traditional ASR system produces a verbatim transcript with 95% accuracy. But the call driver classification incorrectly tags the call as a “general enquiry” rather than a “billing dispute.” The WER looks good. The business insight is wrong. A single WER figure cannot distinguish between these two calls, even though one delivered the right outcome and the other did not.

Under traditional ASR, a 15% WER on a specific product category could be reduced to 8% through a tuning cycle. Under a neural model, the WER on that same category is determined by the model’s training and the audio conditions on that call. Setting WER reduction targets as the main lever for a neural deployment is not just unproductive, it is actively misleading the management audience when it should focus on the real quality lever (the post-processing and summarization layer) which was the whole point of call management with quality resolution.

Part Four: What to Measure Instead — Context Accuracy

The Right Frame: Did the System Understand the Call?

The meaningful quality question for AI-powered contact center transcription is not “how many words did it get right?” but “how accurately did it capture what the call was about?”

This concept — sometimes called comprehension accuracy — is what traditional analytics platforms have long claimed to measure at 90%+. The principle is sound regardless of vendor: the question to ask of any system is whether the output (transcript, summary, classification, sentiment score) correctly reflects the intent, issue, and resolution of the interaction.

For a neural transcription pipeline feeding a post-call summary, this translates into a Context Summary Accuracy (CAS) framework with three measurable dimensions:

1. Call Driver Accuracy Did the summary correctly identify why the customer called? This is the most critical dimension. A call about a failed direct debit that is summarized as a general account enquiry is a meaningful failure regardless of how many words were transcribed correctly.

2. Resolution Accuracy Did the summary correctly capture how (or whether) the issue was resolved? Was an escalation noted? Was a callback promised? Was the customer told to call back? These are the details that drive agent coaching, FCR measurement, and quality scoring.

3. Sentiment Fidelity Did the summary correctly reflect the emotional tone of the call? A frustrated customer who ended the call satisfied should not be logged as a neutral interaction. A customer who appeared satisfied but expressed subtle dissatisfaction about wait times should have that captured.

Why Call Driver Accuracy Cannot Be Read Off a Single Line

The “alright versus right” example in Part Three shows that a transcript can be accurate at the word level and still support a correct summary. The reverse case matters just as much for call driver identification: a transcript can be word-perfect and still not contain the customer’s actual issue in any single sentence.

Customers do not always arrive able to state their problem in the right words. On a complex enquiry, the opening statement is often an approximation, sometimes an inaccurate one, and the real issue only emerges after several rounds of probing, clarifying questions, and restatement from the agent. The call driver, in these cases, lives in the arc of the exchange rather than in any one utterance.

This has a direct consequence for how Call Driver Accuracy should be assessed and, further upstream, for how much weight WER deserves in the first place. If the analysis method is effectively transcript-literal (keyword spotting, or topic modelling that treats the transcript as a bag of words), then transcription accuracy carries the whole signal, and a degraded transcript genuinely limits what can be inferred. But if the method is a model reasoning over the full conversation to infer intent, which is what the CAS framework assumes, transcription quality is a floor condition rather than the main driver: the transcript needs to be clean enough that the exchange is followable, but improving WER beyond that floor does little for call driver identification, while the model’s ability to track a multi-turn clarification pattern does a great deal.

Conclusion

The contact center industry is in the middle of a one-way transition from traditional modular ASR to end-to-end neural transcription. The major platform providers (Microsoft, Google, NICE, Genesys) have all retired or are retiring their traditional ASR engines. New cloud deployments default to neural models. The question is no longer whether to make this transition; for most contact centers, it is already made or imminent.

What has not kept pace is the quality management framework. WER made sense as the primary measure for traditional ASR because errors were systematic, tunable, and measurable against a stable reference. Most of those properties do not hold for neural transcription. WER figures for neural models sit within the same range as traditional ASR on contact center audio, but they cannot be reduced through customer action, and attempts to force the issue through vocabulary boosting or narrow fine-tuning risk trading one class of error for another. Relying on WER alone or treating word-level corrections as a quality program in themselves, creates a false impression of quality management activity.

The replacement framework is Context Summary Accuracy: did the system correctly capture call driver, resolution, and sentiment? This sits alongside WER rather than discarding it: WER remains a useful floor check on transcript quality, but CAS is what can be monitored, tracked, and improved as the primary measure over time. It corresponds directly to the quality lever that neural deployments actually expose: the summarization and post-processing prompt layer. Improving CAS means improving prompts. Improving prompts means improving the business outcomes the contact center cares about: coaching accuracy, FCR measurement, compliance coverage, and Voice of Customer (VoC) insight.

In a neural transcription deployment, WER alone cannot answer the question that matters. The fuller question is whether the system correctly understood the call, and that is a question with an answer you can act on.

References and benchmark sources (see note on sources above): OpenAI Whisper technical report; Nuance Communications product end-of-life notices (2025–2026); Voicegain 2025 STT Accuracy Benchmark (8 kHz contact center audio); VexaScribe / HuggingFace Open ASR Leaderboard data, June 2026; Deepgram “Speech Recognition Accuracy: Production Metrics and Optimisation,” 2025.