Hallucination

What it is

Language-model hallucination is the familiar kind: the agent states an opening time, a policy or a confirmation number that does not exist, in the same steady voice it uses for true statements. Speech-recognition hallucination is less known and specific to voice: given audio that contains no speech (silence, hold music, line noise, a breath), some recognisers emit plausible words. The agent then responds to a sentence nobody said.

Both are worse over the phone than in chat because there is no screen to check and the caller cannot scroll back.

Why it matters when buying

Fabricated transcripts on silence turn every pause into a possible false intent: the caller stops to find their card and the agent hears "cancel my appointment". Fabricated facts create liability, especially in regulated calls (a wrong balance, a wrong due date). Both risks are managed by design: the agent should say "I do not know" when retrieval finds nothing, confirm before any write, and treat empty or very low-confidence turns as silence rather than speech. One practical detector for phantom transcripts is a word timestamp that lies beyond the length of the audio; the methodology page describes the probe.

What to ask

Ask what the agent does with a turn of pure silence or noise; test it by saying nothing for ten seconds. Ask a question the knowledge base cannot answer and listen for invention. Ask what confidence thresholds exist on the speech-recognition side and whether low-confidence words are confirmed before they are acted on.