mirror config.py

config.py

@@ -0,0 +1,54 @@
+"""glossolalia: Coherence Dial configuration."""
+
+from pathlib import Path
+
+PROJECT_ROOT = Path(__file__).resolve().parent
+DATA_ROOT = PROJECT_ROOT / "data"
+ADAPTERS_ROOT = PROJECT_ROOT / "adapters"
+STATIC_ROOT = PROJECT_ROOT / "static"
+
+# TTS base + adapter
+TTS_MODEL_HF = "SWivid/F5-TTS"       # F5-TTS, ~336M, voice-cloning + IPA input
+TTS_MODEL_VARIANT = "F5TTS_v1_Base"
+
+LORA_CONFIG = {
+    "r": 16,
+    "lora_alpha": 16,
+    "lora_dropout": 0.05,
+    "target_modules": ["to_q", "to_k", "to_v", "to_out.0"],
+}
+
+# Dial config: control token = "<CONTROL_STEM> <LEVEL_WORDS[level]>"
+CONTROL_STEM = "tongues"
+LEVEL_WORDS = ["zero", "one", "two", "three", "four"]
+LEVEL_P = [0.0, 0.25, 0.50, 0.75, 1.0]   # phoneme substitution probability per level
+
+# Voice presets (reference clips for F5-TTS zero-shot cloning at inference).
+# Names describe perceived voice character; f0 / accent / license noted for provenance.
+# All clips are CC-BY-4.0, MIT, or Public Domain. Safe to redistribute on a public HF Space.
+# See DECISIONS.md 2026-06-12 voice-preset entry for license sources.
+VOICE_PRESETS = {
+    "v1": {"name": "warm, calm",          "wav": "data/voices/v1.wav", "ref_text": "data/voices/v1.txt"},
+    "v2": {"name": "high and arch",       "wav": "data/voices/v2.wav", "ref_text": "data/voices/v2.txt"},
+    "v3": {"name": "deep, slow",          "wav": "data/voices/v3.wav", "ref_text": "data/voices/v3.txt"},
+    "v4": {"name": "theatrical",          "wav": "data/voices/v4.wav", "ref_text": "data/voices/v4.txt"},
+    "v5": {"name": "tense, anxious",      "wav": "data/voices/v5.wav", "ref_text": "data/voices/v5.txt"},
+    "v6": {"name": "haunted letter",      "wav": "data/voices/v6.wav", "ref_text": "data/voices/v6.txt"},
+    "v7": {"name": "gentle storyteller",  "wav": "data/voices/v7.wav", "ref_text": "data/voices/v7.txt"},
+    "v8": {"name": "fevered poet",        "wav": "data/voices/v8.wav", "ref_text": "data/voices/v8.txt"},
+    "v9": {"name": "stately reader",      "wav": "data/voices/v9.wav", "ref_text": "data/voices/v9.txt"},
+}
+
+# Validation gates
+WHISPER_MODEL = "base.en"
+RESEMBLYZER_MIN_COSINE = 0.85
+SPEARMAN_GATE = 0.80                 # WER vs level Spearman; positive (WER rises as dial rises)
+
+# Audio
+SAMPLE_RATE = 24000                  # F5-TTS native
+GEN_MAX_SEC = 30
+
+# HF repos
+HF_DATA_REPO = "akshan-main/glossolalia-inputs"
+HF_LORA_REPO = "akshan-main/glossolalia-dial-lora"
+HF_SPACE_REPO = "akshan-main/glossolalia"