NPC voice bank, build 24 / 2026-09-14
343 locally generated clips. These are synthetic performances, not human actor recordings or voice clones.
Model: hexgrad/Kokoro-82M v1.0 (Apache-2.0); https://huggingface.co/hexgrad/Kokoro-82M
Model SHA256: 496dba118d1a58f5f3db2efc88dbdc216e0483fc89fe6e47ee1f2c53f18ad1e4
Inference library: kokoro 0.9.4; https://github.com/hexgrad/kokoro
License text: Kokoro-LICENSE.txt.
Rendered on this Mac using MPS. No original player recordings were uploaded, cloned or changed.
Dialogue: existing realHobo character text and new flee/cover/clear reactions.
manifest.json records the exact text, profile, speed, duration, filename and SHA256 of every clip.
17 character delivery profiles use 9 underlying English voice presets. Characters sharing a voice retain different dialogue and pacing.
Master: 24 kHz PCM. Game copies: mono MP3, 112 kbps, 70 Hz high-pass, -18 LUFS normalization, -2 dB true-peak target. No pitch shifting or radio effects applied to characters.
The user's player and intercom recordings remain separate assets.
Regenerate: Godot --headless --path game --script res://scripts/export_npc_voice_plan.gd, then tools/build_npc_voices_v24.py in the documented Python runtime.
Only audio and this manifest are required at runtime. The game does not use system TTS, download models, or contact a speech service.