The "Voice Browser " activity within the W3C consortium addresses the need for standards for speech grammars, dialogue descriptions etc. in distributed systems. This paper discusses the consortium's recent speech grammar working draft specification. The W3C specification is based on the Javatm Speech Grammar Format (JSGF) defined by Sun Microsystems and is - with all good and bad qualities - characterized by traditions of formal language theory. In a constructive spirit, we suggest some possible improvements based on natural language theory. The suggestions concern compound feature-based semantic presentations, lexicon structure, ambiguity, and exhaustive parsing,