config.yaml

# UsTwo Project Configuration
project:
  name: "UsTwo"
  version: "0.1.0"

paths:
  data_dir: "data"
  samples_dir: "data/samples"
  models_dir: "data/models"
  output_dir: "data"

# Stage 1: Speaker Diarization + ASR
stage1:
  output_path: "data/stage1_output.json"
  segments_dir: "data/segments"

  preprocessing:
    target_sample_rate: 16000
    max_duration_sec: 300
    min_duration_sec: 3
    target_peak: 0.95  # peak normalization — handles volume differences across devices

  diarization:
    model: "pyannote/speaker-diarization-3.1"
    num_speakers: 2
    merge_gap_sec: 0.15

  asr:
    model: "large-v3-turbo"
    compute_type: "int8"
    language: null  # null = auto-detect
    batch_size: 16

  alignment:
    enabled: true

  language_id:
    enabled: true
    # SenseVoice disabled — Whisper language + text heuristic 사용
    # emotion2vec finetuning 실패 시 SenseVoice 감정 힌트 활용 예정
    # model: "FunAudioLLM/SenseVoiceSmall"

# Stage 2: Audio Emotion + Text Emotion
stage2:
  input_path: "data/stage1_output.json"
  output_path: "data/stage2_output.json"
  audio_emotion:
    model: "iic/emotion2vec_plus_base"
    lora_onnx_path: "data/models/lora_emotion2vec_7class/model.onnx"
    finetuned_checkpoint: null  # legacy, use lora_onnx_path instead
  text_emotion:
    korean_model: "searle-j/kote_for_easygoing_people"
    korean_lora_onnx_path: "data/models/lora_kcelectra_7class/model.onnx"
    korean_lora_tokenizer: "data/models/lora_kcelectra_7class/best_model"
    english_model: "j-hartmann/emotion-english-distilroberta-base"
  fusion:
    mode: "emotion_specific"  # "emotion_specific" (per-class grid-search optimized) or "fixed" (60/40)
    audio_weight: 0.6  # fallback for mode="fixed"
    text_weight: 0.4

# Stage 3: Character Reaction + Garden + Recap
stage3:
  input_path: "data/stage2_output.json"
  output_path: "data/stage3_output.json"
  recap:
    llm_provider: "anthropic"   # anthropic | openai
    model: "claude-sonnet-4-20250514"
    max_tokens: 500
  garden:
    growth_per_call: 5
    max_level: 5

# Stage 4: FastAPI Server
stage4:
  input_path: "data/stage3_output.json"
  host: "0.0.0.0"
  port: 8000
  max_upload_size_mb: 50
  allowed_extensions: [".wav", ".mp3", ".m4a", ".ogg"]

# API Keys (환경 변수에서 로드 — .env 파일 사용)
api: {}

# Logging
logging:
  level: "INFO"
  format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"