How does automatic speech recognition work?
An ASR system takes audio from a microphone or a recording and outputs the most likely sequence of words. Older systems combined separate acoustic, pronunciation and language models, while most current ones are neural networks trained on large amounts of transcribed speech.
It runs in two modes. Batch transcription processes a finished recording. Streaming recognition returns words while the person is still speaking: the Web Speech API, for example, gives a page interim results, which can still change, followed by final ones.
Accuracy is usually measured as word error rate (WER), which Google Cloud’s documentation calls the industry standard. Count the inserted, substituted and deleted words against a correct reference transcript, then divide by the number of words in the reference. A WER of 5% means about one error for every twenty words.
Example: a caller says “I can’t see the billing tab”. A streaming recogniser may show “I can’t see the” first, then settle on the full sentence once the caller pauses.
Why does speech recognition matter for a voice agent?
A voice agent can only answer the question it heard. Product names, error codes, email addresses and accents are where recognition tends to slip, and one misheard word can lead to the wrong answer. OpenAI’s guidance on evaluating voice agents says to test recognition across accents, background noise, language switches, names and numbers.
Recognition matters after the call too, because the transcript is how a team reviews what customers asked and finds gaps in its documentation.
ASR works out what was said, not who said it. Recognising a person from their voice is speaker recognition, a different technology that treats the voice as biometric data.
How does Kawl Now handle speech recognition?
Kawl Now doesn’t run a separate speech-to-text step that you set up. The caller’s audio goes to OpenAI’s GPT-Live voice model, which listens to it directly and also returns a running text transcript of both speakers.
The widget can show that transcript as live captions, the last eight turns at a time, and Kawl Now saves it as the conversation’s two-sided transcript. When you test, read the caller’s lines first: if they don’t match what was said, check the microphone and background noise before you change your knowledge sources.
The browser is asked to apply echo cancellation and noise suppression to the microphone. There’s no language setting or custom vocabulary list, and agents greet callers in English.
Common questions
Is speech recognition the same as voice recognition?
Not quite. Speech recognition works out what was said. Voice recognition often means speaker recognition, which works out who is speaking from the sound of their voice. Some products do both, which is why the terms get mixed up.
What is the difference between ASR and speech to text?
There isn’t one in everyday use: speech to text is the plain-English name for automatic speech recognition. Vendors often call the product speech to text and the underlying technology ASR.