sine2pi/ASR-model
ASR model
What's novel
ASR model
Code Analysis
6 files read · 3 roundsAn experimental automatic speech recognition (ASR) system for LibriSpeech that combines a custom convolutional audio encoder, novel Gumbel-routed normalization, gammatone-scaled rotary attention, pitch-aware value gating, and a custom row/column-variance optimizer with max/median scaling strategies.
Strengths
Genuinely novel components including a custom optimizer with depth-based role assignment and pitch/spectrogram-aware scaling, a Gumbel-softmax routed normalization layer, and gammatone-frequency-scaled rotary embeddings. The code contains real algorithmic depth rather than thin wrappers around existing libraries.
Weaknesses
No tests, no documentation (empty README), unconventional naming (using 'n' for self), extensive commented-out experimental code, and several potential bugs (e.g., attention forward has unreachable code paths, variable shadowing in loops). The architecture is a tangled research prototype rather than a maintainable system.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
590
Contributors
1
Files
6
Active weeks
32
Repository
Language
Python
Stars
1
Forks
0
License
—