What is speech to text: STT explained
Speech to text is the automated conversion of spoken audio into written words, producing a machine-readable transcript that can be searched, quoted, and stored next to the recording it came from. It is also called STT, automatic speech recognition, or ASR, and most product documentation treats the three terms as the same thing.
The distinction worth holding onto is between the process and the artefact. Speech to text is the machine process. Transcription is the written record it produces, and that word still carries its older meaning of a person doing the same job by ear.
How does speech to text work?
The audio is sliced into short frames, typically tens of milliseconds each. An acoustic model maps each frame to sub-word sound units, and a language model scores the candidate word sequences those units could form, preferring sequences that are plausible in the target language.
Two consequences follow from that second step, and both catch people out:
- The system guesses in context. A word that was never spoken can appear in the transcript because it fits the surrounding sentence better than what was actually said.
- Rare words lose. Part numbers, model codes, surnames and local place names are the tokens a language model has seen least often, so they are the tokens it most readily overwrites with something common.
Try Venta Capture on your own process
Build one flow for your highest-volume case type. Free, no credit card.
How is speech to text accuracy measured?
The standard measure is word error rate, usually written WER. NIST publishes the reference scoring toolkit, SCTK, whose sclite tool aligns machine output against a human reference transcript and counts three kinds of error: substitutions (S), deletions (D) and insertions (I). WER is (S + D + I) divided by N, the number of words in the reference.
Microsoft's Azure documentation works a small example of exactly this arithmetic. A 13-word reference sentence in which one word is substituted, one is dropped, and one word is wrongly split into three scores 5 divided by 13, a WER of 38 percent. Worth noting: because insertions are counted but the denominator is fixed, WER can exceed 100 percent.
Where speech to text is weakest
Published accuracy figures almost always come from clean, close-microphone, read speech. Audio from a phone held at arm's length in a workshop or on a roadside is a different problem, and the gap is not small.
- Accent and dialect. Koenecke and colleagues, writing in PNAS in 2020, tested five commercial ASR systems from Amazon, Apple, Google, IBM and Microsoft and found an average WER of 0.35 for Black speakers against 0.19 for white speakers. Same systems, same task, roughly double the error rate.
- Background noise and reverberation. Compressors, traffic, wind and hard-walled rooms all push WER up.
- Overlapping speech. Two people talking at once degrades output badly unless the system does explicit speaker diarisation.
- Domain vocabulary. Terms of art, abbreviations and identifiers are the highest-value words in an operational recording and the least reliably transcribed ones.
Speech to text explained: a practical example
A driver records a fault and says: "there is a knock from the offside front, VIN ends 4471." The transcript comes back as "there is a knock from the outside front, in ends four four seven one." Every word is plausible English and the sentence reads perfectly well. Two of the three facts an operator needs are now wrong, and nothing in the transcript marks them as suspect.
That is the characteristic failure of STT. It rarely produces gibberish you would catch at a glance. It produces fluent text with quiet substitutions inside it.
What speech to text does not do
It does not understand. The output is a word sequence, not an interpretation, not a decision, and not a summary of what mattered in the recording. It also does not verify anything: a confident transcript of a false statement is still a transcript of a false statement.
The practical stance is to treat the transcript as a search index over the audio rather than a replacement for it. Keep the original recording, let a reviewer jump to the timestamp, and never let a downstream rule fire on a transcribed identifier without a human or a second signal confirming it. Pairing spoken explanation with a photographed label and optical character recognition of that label is more reliable than a spoken number alone, which is why both usually appear together in claim documentation workflows.
Try Venta Capture on your own process
Build one flow for your highest-volume case type. Free, no credit card.