AGENTS.md

Speech Processing AGENTS

Purpose

PyQt6 desktop app for speech recording, segmentation, feature extraction, and recognition with DTW/HMM backends plus Kalman speech-following mode.

File Map

• UI and mode routing: speech_recognition_app.py
• DSP and feature extraction: speech_dsp.py
• Recognition backends: speech_hmm.py
• Dictionary persistence/schema helpers: dictionary_store.py
• Audio I/O adapter: audio_io.py
• Dictionary data: speech_dictionary.json

Run

• python3 speech_recognition_app.py

Current Workflow Decisions

• Mode switch is menu-driven:
  • Speech Recognition
  • Kalman Speech Following
• Recognition backends supported in UI:
  • dtw
  • hmm-discrete
  • hmm-continuous
• Segment selection strategy is explicit:
  • manual
  • auto-best (default for save/compare automation flows)
• Advanced DSP parameters are shown in a collapsible panel; keep compact startup behavior.
• Presets are part of the normal flow (Quiet Room, Noisy Room, Fast Speech, Custom).

Defaults And Ranges

• Sampling rate: 8000 Hz
• Frame length: 30 ms (recommended range 20-40 ms)
• Hop length: 10 ms (recommended range 5-15 ms)
• Pre-emphasis: 0.97 (recommended range 0.95-0.98)
• Threshold ratios (energy, zcr, entropy): typical 0.05-0.20
• LPC order: 12 at 8 kHz (recommended 10-14)
• Wavelet scales: 6-24

Kalman Mode Notes

• Supports offline source modes (auto-best segment, manual segment, whole recording) and live microphone tracking.
• Live mode confidence is based on rolling innovation stability; keep this metric visible in UI updates.

Data Compatibility Notes

• Keep backward compatibility with existing entries in speech_dictionary.json (including legacy MFCC-only records).
• New saves should continue to include extraction settings and selected feature matrices alongside spectrogram metadata.