manifest: declare the audio frontend, transducer decode and driver
#1
by Void2377 - opened
The published manifest lists the two graph packages and nothing else, so a runtime cannot drive this bundle: no mel settings, no blank id, no duration table, no predictor state names.
This is the same file regenerated from the recipe and the COMPILED graphs, adding audio, decode and a one-step plan naming the transducer driver. The graph package filenames are unchanged, so no weights move.
Verified: with this manifest the bundle transcribes LibriSpeech 1272-128104-0000 at word error rate 0.0000, both through the NeuRT runtime directly and through the RunAnywhere SDK's speech interface.
Pairs with RunanywhereAI/neurun#60 and RunanywhereAI/runanywhere-sdks#710.
sanmonga22 changed pull request status to merged