Talking Clock Project

Live demo here.

Goal

Our goal is to build a talking clock with the following specifications :

• Clients sole responsability is to retrieve and play audio files
• A server acts as the time reference, build on-the-fly audio files and triggers 'play events'
• Use of an external text2speech solution is disallowed (otherwise eSpeak will kindly do the job)
• Homemade talks are composed thanks to the minimum set of personal audio patterns

Design choices

• HTML5 Audio tag for broad client support
• Default use of .ogg audio file (fallback to .mp3 if client requires)
• Node.js, socket.io and node-schedule for sync management from server side

Preliminary treatments

1. We record a minimum list of elementary voice patterns "0, 1, et 1, 2:16, 20, 30, 40, 50, heure, minute, seconde" using an iPhone voice recorder
2. Audacity "Sound Finder" & "Export Multiple” functions are then used to isolate and split patterns from previous record
3. Finally, a Python script uses Sox to recompose an exhaustive list of spoken numbers ranging from 0 to 59

Workflow

1. The server emits regularly a 'play event' with respect to a 5 seconds node-schedule
2. Clients retrieve the current audio file and start playing
3. During that time, the server prepares the next file to play : Sox composes on-the-fly audio sentences (join & merge elementary audio patterns) and persists corresponding ogg and mp3 files

Concluding remarks

• An other tempting approach is to run a streaming server as IceCast feeded by ezStream it-self plugged on a dynamically generated list of audio files. This approach is misleading : managing time sync on server side (both at inception and periodic resets) is tricky in this framework
• Finally, a valid (but cumbersome) way is to use a WebRTC framework. It's over optimized for our needs and we haven't tested it
• Last word : our solution doesn't take into account client internet bandwidth. UX will be poor if it's too low !