Keyword Spotting and Full Transcription

Speech analytics products are built on one of two approaches, and vendors are not always explicit about which.

Phonetic keyword spotting

Searches the audio for sound patterns matching specified terms, without producing a transcript.

Advantages: fast, computationally cheap, robust to some vocabulary problems because it matches sounds rather than words, and it handles novel terms without retraining.

Limitations: you can only find what you configured. No transcript means no summarisation, no classification of unanticipated topics, and no ability to ask a question you did not plan.

Where it survives: high-volume compliance checking on a fixed term list, and as a fast pre-filter.

Full transcription

Produces text for every call, which is then searched, classified and analysed.

Advantages: the transcript is an asset. Any question can be asked of it later, including questions nobody anticipated. It supports summarisation, categorisation, retrieval and everything built on language models.

Limitations: more expensive, slower, and its accuracy varies by speaker in the ways described elsewhere.

This is what modern platforms do, and the cost has fallen enough that the argument is largely settled.

The decisive difference

With keyword spotting, a new question requires reconfiguring and reprocessing — if the audio is still retained.

With transcription, a new question is a search over existing text.

When a regulator asks whether a particular thing was ever said, or when a product problem emerges and you need to know when customers started mentioning it, this difference is the whole answer.

The transcript is the durable asset. Categories, scores and dashboards are derived and replaceable; the text is not, and losing the ability to re-ask questions is the real cost of the cheaper architecture.

The hybrid position

Several platforms use both.

Keyword spotting as a fast filter, transcription on the calls that matter.

Reasonable where volume makes full transcription costly and less common now that costs have fallen.

Ask which is used for which function, because a compliance check running on phonetic matching and a summarisation running on transcription have different accuracy characteristics and the interface will not distinguish them.

What to check either way

Precision and recall on your own terms. Take a term that matters, find fifty calls the system says contain it, and check. Then find fifty it says do not.

Performance on your product vocabulary, which is where both approaches struggle.

Behaviour across accents, which differs between the architectures and should be measured rather than assumed.

What happens to the audio afterwards. If the architecture depends on reprocessing, retention of the audio becomes a functional requirement as well as a legal question.

The practical recommendation

Prefer full transcription unless volume economics genuinely rule it out.

Retain the transcript on the retention schedule the analytics purpose justifies, which may differ from the audio's.

Treat the transcript as the asset and build the categories on top of it, so that changing category definitions does not require reprocessing audio you may no longer hold.

And check whether the transcript is exportable, because a platform holding the only copy of your text has the same lock-in problem as one holding the only copy of your history.

The reprocessing question

A specific question to settle during procurement, because it determines what you can do later.

If a category definition changes, is historical data reclassified? With a transcript, yes. With keyword spotting, only if the audio is retained and reprocessed.

How far back can reprocessing run, and at what cost?

Is the transcript retained independently of the audio, so that audio deletion does not destroy analytical capability?

Can transcripts be exported and reanalysed outside the platform?

What happens to historical determinations when a model is updated — restated, preserved, or both?

The answers determine whether your archive is an asset or a snapshot. An operation that discovers after two years that it cannot re-ask a question of last year's calls has lost something it did not know it was choosing to give up.