The L2 link is between the two 1812 routers? What type of physical link is it and what's its physical bandwidth?
If the link is physically 10 Mbps, you shouldn't need to use a shaper. If it's physically faster and policed by a provider, you might need to account for what time intervals the provider uses and/or shape slower than the nomimal rate to account for L2 overhead.
What traffic starts to see 20 ms response times? If normal data traffic, as the link approaches 100%, traffic flows should start to queue. What your current policy should provide, is all flows, except the LLQ voice, should slow, more or less, equally.
QoS doesn't elminate the need for bandwidth, it just allows you to manage it. So, when the link is congested, you can insure some traffic gets the bandwidth, e.g. voice, but at the expense of less bandwidth provided to other traffic. If you had some large bulk flows running with your other "normal" data traffic, if you can identify that traffic, you could pull it out of the default class and place it into it's own class with a minimum bandwidth setting. This should improve the performance for the traffic left within the class-default FQ but at the expense of the extracted bulk traffic (usually an acceptable trade-off).
[edit]
PS:
As a rule of thumb, to avoid queuing for interactive type traffic, average bandwidth utilization below 33.3% is excellent. Up to 66.7% is good to acceptable. Above 66.7%, queuing can be become significant. So, you're issue above 80% is likely normal. As noted above, you might be able to further manage your traffic, or you'll need more bandwidth.
