Techniques are disclosed for producing high quality losslessly compressed audio tracks based on conversations between participants remote from one another, such as conversations that occur during a telephonic interview or online conference, or other conversations that take place over a network between two or more participants. In an embodiment, each participant's device includes an audio chat client configured to record that participant's audio contribution to the conversation and store a non-compressed version of the contribution locally. A first version of the captured audio is generated with lossy compression and pushed in real time to a cloud-based service, for purposes of the live conversation. A second version of the captured audio for subsequent playback is generated and stored with lossless compression and is pushed asynchronously to the service. The service is configured to automatically provide a multitrack project with high quality audio tracks from each participant.