I bought a couple of the MMA7660 accelerometers I mentioned some time ago from Farnell (before their gender op). The chips are incredibly tiny, 3mm to a side, and 0.5mm pitch legless contacts have to be seen to be believed. Soldering wires onto those was beyond me, and I was fortunate to have Aaron at NICTA kindly solder one to a board for me. Etienne more-or-less inhaled the other one. Some basic prodding seemed to indicate that the device was still functional after this surgery.
Much later I tried to solder some wires to the board, using NICTA's excellent facilities (Dremel, high-end electric iron and stereoscopic magnifier). Suffice it to say that the device showed no signs of life after that. Fortunately the market has responded to my demand in the intervening period: the Mad Scientist Hut sold me a couple of these devices pre-attached to macro breakout boards. Their prices are OK but their shipping is very expensive, at about $10 for the pair. I guess it would make sense if I'd bought fifty of them, but I didn't.
Suffice it to say that the bidirectional circuit mentioned in SparkFun's tutorial did the trick of interfacing the circa 3v TWI/I²C levels of the accelerometer to the circa 5v used by the AVR and the rest of the circuit. I fabbed it on some stripboard and get plausible readings from the sensor.
The penultimate bit of hardware hacking was to switch off the speech chip when quiescent, with the goal of getting the current draw under a milliamp in the most-of-the-time case. Doing this with is a FET is entirely straightforward, but the current draw remains a ridiculous 6.5mA. I think (hope) that is fixable in software. Adding the batteries — a button cell as a backup for the RTC, and four NiMH AAs — is largely a mechanical problem. Later I might also try to bring the volume under software control.
I have begun trying to flesh out the control software for this thing. I'm trying to avoid writing spaghetti C and have been a little successful, but am hoping for a more abstract way of writing the core state machine as it will involve commands coming as data on the U(S)ART and timeouts, as well as the presumably complicated interactions the accelerometer will allow. Maybe I can get by with less overkill than Esterel.
The code is at github. Possibly of interest to others is the growing MMA7660 driver for AVR.