What is speech to text?
Speech to text is the conversion of spoken audio into written text, which is the same task the literature calls automatic speech recognition. The two names describe identical work, with speech to text being the term product teams use and procure against. What varies between implementations is streaming behaviour and the handling of telephone audio rather than the definition.
Also called: Automatic speech recognition, Transcription
Where two names for one task cause real procurement confusion
Vendors describe the same capability as speech to text and as automatic speech recognition, and buyers reasonably assume the two are different products. They are one task, so a comparison has to rest on the characteristics that genuinely differ between implementations. Whether a model returns partial results while the caller is still speaking matters far more to a voice agent than the wording on a datasheet. So does whether accuracy was measured on clean read speech or on eight kilohertz telephone audio with a caller standing in traffic.
- Speech to text and automatic speech recognition name the same task exactly.
- Streaming behaviour differs between vendors far more than headline accuracy does.
- Accuracy quoted on clean audio has little bearing on telephone performance.
- Indic language support varies sharply between providers of the same capability.
How speech to text works
- Audio arrives as a stream and is divided into short analysis frames.
- An acoustic model produces a distribution over sound units for each frame.
- A decoder combines those with a language model to produce candidate text.
- Partial results are emitted while the caller is still speaking.
- A final transcript is committed for the utterance once speech ends.
- Domain vocabulary is supplied as a bias list to improve rare term accuracy.
How speech to text is evaluated
Speech to text is evaluated on audio matching the deployment rather than on a benchmark corpus. Measure the following on each language the agent supports:
- word error rate on telephone audio from the live carrier path
- accuracy on domain vocabulary and proper nouns
- latency from end of speech to a committed transcript
- stability of partial results as an utterance proceeds
- degradation under background noise at realistic levels
Limitations, risks and common failure modes
- Benchmark figures are quoted from read speech and never reproduced on calls.
- Partial results that revise themselves confuse logic acting on them too early.
- Code mixed utterances are transcribed into one language and lose half the content.
- A bias list grows large enough to introduce errors of its own.