Skip to content

claude.md

Latest commit

 

History

History
53 lines (36 loc) · 1.55 KB

claude.md

File metadata and controls

53 lines (36 loc) · 1.55 KB

Claude Code Assistant Guidelines

Git Policy

NEVER commit changes - User handles all git operations. You may use git status, git diff, git log for inspection only.

Build After Changes

Always build immediately after code changes:

cmake --build build --config Release --target parakeet_cli

Quick Reference

Executable: build/examples/cpp/Release/parakeet_cli.exe Model Cache: %LOCALAPPDATA%\eddy\cache\models\parakeet-v2\ Performance: 5-8x RTF on CPU, 1.27% WER on LibriSpeech

Project Goal

Match FluidAudio's Parakeet v2 (Swift/CoreML) implementation in C++/OpenVINO:

  • ✅ 4-Model Pipeline, Greedy Decoding, LSTM State Continuity
  • ✅ Token Deduplication, Timestamps, Confidence Scores
  • ⏳ Decoder Output Caching (next priority)

Current Results: 1.27% WER (better than FluidAudio's 2.2% baseline)

Architecture

Batch Chunking (current focus):

  • 10s chunks with 3s overlap
  • 2D search deduplication at boundaries
  • LSTM state continuity across chunks

NOT building streaming yet - focus is on batch processing of complete audio files.

Testing

# Quick test
build/examples/cpp/Release/parakeet_cli.exe "assets/audio/first_15s.wav"

# Benchmark (C++ native, 2.5x faster than Python)
build/examples/cpp/Release/benchmark_librispeech.exe --max-files 25

Key Implementation Notes

  • Release builds only (Debug requires VS runtime DLLs)
  • Token timestamps: frame × 0.08 = seconds
  • Confidence: softmax over joint network logits
  • Deduplication: checks last 20 tokens vs first 15 tokens of next chunk