Building Speech Analytics Rather Than Buying
Open transcription models have made the technical build tractable. Whether to do it turns on the operational layer, which is most of the product.
Speech recognition of usable quality is now available as open models that run on your own hardware. That changes the build question genuinely, and less than it appears.
What building involves
Transcription. An open model, self-hosted. Quality on telephony audio is respectable and improving, and it can be adapted to your vocabulary.
Storage and indexing. Text search over transcripts.
Categorisation. Rules, a classifier, or a language model.
Acoustic measures, which are straightforward signal processing.
A user interface, which is where the effort actually goes.
The transcription is the part people assume is hard and it is the part that is solved.
What building buys
Data stays in your estate. For regulated sectors, or where the vendor terms on training are unacceptable, this is frequently the whole argument.
No per-minute pricing, which at scale is substantial.
Full control of categories and definitions.
Model version stability. Nothing changes unless you change it, which matters for compliance reproducibility.
Custom vocabulary and adaptation on your own audio without asking a vendor.
What it costs
The operational layer. Search interface, evaluation workflow, scorecard management, calibration tooling, dispute handling, reporting, access control, audit logging. This is most of a commercial product and none of it is the interesting part.
Ongoing engineering. Models improve, audio infrastructure changes, requirements grow.
Accuracy work. Vocabulary, adaptation, measurement by speaker group. A vendor does some of this for you.
Key person risk. A system understood by one engineer.
No vendor to hold accountable when it is wrong.
The hybrid that usually wins
Self-host transcription, which is the part with the data sensitivity and the per-minute cost.
Buy or build lightly on top for search and analysis.
Keep the evaluation workflow in whatever your operation already uses, rather than building a QA application.
This gets the data control and the cost saving without rebuilding a product, and it is the arrangement most organisations with an engineering capability end up at.
When to build
Regulated data that cannot leave the estate.
Volume where per-minute pricing is material.
A requirement no vendor meets — an unusual language, an unusual integration, a specific compliance need.
Existing engineering capacity with a named owner for three years.
Reproducibility requirements that a vendor's model update cycle cannot satisfy.
When not to
When nobody will own it. An unmaintained pipeline degrades quietly and the degradation is invisible.
When you need the workflow layer, which is where the effort is and where a vendor genuinely adds value.
When the operation is small. The fixed cost of building does not amortise.
When the driver is cost alone and the volume does not justify it.
The honest test
Who owns this in three years? A name and allocated time, or it will decay.
What are we actually buying from a vendor? If the answer is transcription, building may be right. If it is the workflow, the scorecard tooling, the calibration and the dispute handling, building means rebuilding all of that.
Most organisations discover the answer is the second, and that the transcription they thought they were paying for is the cheapest part.
A minimum viable build
For a team that has decided to build, the smallest thing that produces value.
Self-hosted transcription on your own audio, with a custom vocabulary.
A transcript store with full-text search.
Acoustic measures computed at ingest — silence, talk-over, talk ratio, monologue length.
A rules engine for categories, with definitions in version control.
Precision measurement tooling: a way to sample matched calls and record whether each was correct.
Export to whatever the operation already uses for evaluation and reporting.
No evaluation application, no dashboard, no scorecard tooling. Those are the expensive parts and they duplicate systems you already have.
This is a few weeks of engineering and it delivers search, categorisation and acoustic analysis, which is most of the operational value of a commercial platform.
More in this section