Skip to content

Getsu v1.2.7

Choose a tag to compare

@skttr87 skttr87 released this 05 Oct 17:31
· 24 commits to main since this release
  • Proactive Host API Resolver: Introduced src/host_api_resolver.py (resolve_pair and resolve_single) to inspect and align duplex input/output endpoints onto the same safe Windows Host API (prioritizing WASAPI -> DirectSound -> MME) before stream initialization, eliminating PortAudio host mismatch crashes (PaErrorCode -9993).
  • Elimination of WDM-KS Driver Crashes: Proactively avoids PortAudio's unstable WDM-KS backend on Realtek onboard audio jacks (GLE 0x492 / ERROR_SET_NOT_FOUND / PaErrorCode -9999), transparently mapping selected endpoints to modern, stable WASAPI equivalents.
  • Intelligent Jack Role & Token Matching: Resolves hardware jack endpoints (e.g., WDM-KS "Mic in at front panel (black)") to matching software endpoints (e.g., WASAPI "Microphone (Realtek(R) Audio)") using acoustic role dictionaries (mic, line, mix) and token similarity.
  • PortAudio Channel Desync Guard: Updates stream channel counts (in_channels, out_channels) directly from resolved device metadata, preventing mono/stereo channel mismatch invalidations (PaErrorCode -9998).
  • Host-Guarded Stream Settings: Enforces sd.WasapiSettings(auto_convert=True) strictly on WASAPI streams, preventing invalid parameter exceptions (PaErrorCode -9984) on DirectSound and MME fallbacks.
  • Synchronized GUI & Voice Test Engine: Updated src/gui.py to route both the live engine and the "Hear Myself" 11s voice test through the resolver. Resolved device indices are safely persisted back to config.json, with visual indicators if endpoint mapping was ambiguous.
  • Energy-Guarded Consonant Sustain (Dual-Condition Keep-Open): Added acoustic energy evaluation (input_rms_db - noise_floor_db >= 6.0 dB and P >= 0.25) capped strictly to 400ms following genuine voiced frames (_frames_since_voiced <= 40). Unvoiced stops (/p/, /t/, /k/) and sibilants/fricatives (/s/, /sh/, /f/) are preserved with 100% fidelity without risk of perpetual open states in continuous ambient noise.
  • Calibrated Near-Field Cooldown Breakthrough: High-energy near-field speech (SNR >= onset_snr_db + 8 dB, P >= 0.55) immediately breaks through post-close cooldown, preventing swallowed syllables while maintaining robust immunity against background chatter and mechanical transient clicks.