The system is designed to interface with external devices and services, to transcribe audio that may be stored elsewhere such as a wireless phone'voice mail, or occurring between two or more parties such as a conference call. An audio stream is separated into many audio shreds, each of which has duration of only a few seconds and cannot reveal the context of the conversation. A workforce of geographically distributed transcription agents who transcribe the audio shreds is able to generate transcription in real time, with many agents working in parallel on a single conversation. No one agent (or group of agents) receives a sufficient number of audio shreds to reconstruct the context of any conversation. The use of human transcribers allows the system to overcome limitations typical of computer-based speech recognition and permits accurate transcription of general-quality speech even in acoustically hostile environments.