Previous research has shown that conceptual/semantic information precedes segmental information in stress language production (such as English or Dutch). The present study investigated the relative time course of semantic and metrical encoding during Chinese (a tone language) production. Participants were shown pictures and carried out a dual-choice go/nogo decision based on semantic information (whether the depicted item was of a living or non-living object) and metrical information (whether the tone of picture's name was one type, tone 1 or tone 2, or the other type, tone 3 or tone 4). The behavior data, the N200 (related to response inhibition) and the lateralized readiness potentials (LRP, related to response preparation) indicated that semantic encoding was prior to metrical encoding. It provided evidence for early semantic encoding and later metrical encoding of speech production, and supported the serial model or the cascading model, rather than the parallel model of language production. This conclusion relating to the time course of semantic versus segmental encoding in stress languages might be generalized to semantic versus metrical encoding in tone languages such as Chinese.