-
Fixed bug in which a very slow 3rd-pass search could result in a leak at the time of stopping recognition, or delay stopping unnecessarily even without a leak (fix affects RapidEars as well).
-
Fixed bug in which a race condition could cause hanging at the time of stopping recognition on very new devices.
-
Fixed bug in which English-language g2p in two different view controllers could cause an exception.
-
Optimization of language model generation time (in some cases notable, and even bigger improvements to RuleORama grammar generation time are forthcoming shortly).
-
Small increase of default vadThreshold.
-
Added the ability to skip setting the sample rate, buffer size, and number of channels while the audio session is set. These are debugging tools and shouldn’t be needed under normal operation, with the small exception that apps which do simultaneous playback of more than 16k sample rates may experience playback quality benefits from disabling preferred sample rate. Some audio session changes to improve coexistence under some circumstances.
-
Improved logging of audio settings.
-
Turned pedantic warnings off since in combination with warnings as errors they were tending to make the project uncompilable for many developers after any LLVM changes that added warnings to pedantic.
-
Turned bitcode off in the sample app.