This is still work in progress, more to come
-
Integrate https://github.com/jcsilva/docker-kaldi-android build scripts and make sure it builds for all architectures (x86, armv7, arm64). Think of moving Kaldi build to cmake for quick portability.
-
Build proper optimized large vocabulary model (maybe 20k words from teldium, 4-5 layers, small enough to run on mobile)
-
Move to grammar decoder to avoid huge HCLG model. https://github.com/jpuigcerver/kaldi-decoders
-
Load model from the AAR (mmap them in tflite style)
-
Add decoding speed measurement
-
Add wakeup word
-
Integrate proper hardware optimized neural network library. Candidates are:
-
Quantization for the models