A method and apparatus are disclosed for validating agreement between textual and spoken representations of words. A voice input verification process monitors a conversation between an agent and a caller to validate the textual entry of the caller's spoken responses or the agent's spoken delivery of a textual script (or both). The audio stream corresponding to the conversation between the agent and the caller is recorded and the textual information that is entered into the workstation by the agent is evaluated. Speech recognition technology is applied to the recent audio stream, to determine if the words that have been entered by the agent can be found in the recent audio stream. The grammar employed by the speech recognizer can be based, for example, on properties of the spoken words or the type of field being populated by the agent. If there is a discrepancy between what was entered by the agent and what was recently spoken by the caller, the agent can be alerted and the error can optionally be corrected.