config.py

from pathlib import Path
ROOT = Path(__file__).resolve().parent
DATABASE_DIR = ROOT / 'database'
AUDIO_DIR = DATABASE_DIR / 'audio'
SAMPLE_CLIPS_DIR = AUDIO_DIR / 'sample_clips'
EVAL_DIR = AUDIO_DIR / 'eval'
OOD_TEST_CLIPS_DIR = EVAL_DIR / 'ood_test_clips'
TRUNCATION_PILOT_DIR = EVAL_DIR / 'truncation_pilot'
LIVE_RUNS_DIR = AUDIO_DIR / 'live_runs'
CHECKPOINTS_DIR = DATABASE_DIR / 'checkpoints'
MODELS_DIR = DATABASE_DIR / 'models'
EASY_TURN_DIR = MODELS_DIR / 'easy_turn'
CACHE_DIR = DATABASE_DIR / 'cache'
DATA_CACHE_DIR = CACHE_DIR / 'dataset'
OOD_CACHE_DIR = CACHE_DIR / 'ood_test'
EXPERIMENTS_DIR = ROOT / 'experiments'
EMBEDDINGS_CACHE_DIR = EXPERIMENTS_DIR / 'embeddings_cache'
RESULTS_DIR = EXPERIMENTS_DIR / 'results'
SAMPLE_RATE = 16000
SMART_TURN_WINDOW_SECONDS = 8
SMART_TURN_WINDOW_SAMPLES = SMART_TURN_WINDOW_SECONDS * SAMPLE_RATE
WHISPER_TINY_ID = 'openai/whisper-tiny'
WHISPER_BASE_ID = 'openai/whisper-base'
WAV2VEC2_ID = 'facebook/wav2vec2-base'
QWEN_LOCAL_ID = 'Qwen/Qwen2.5-0.5B-Instruct'
HINDI_CTC_ID = 'theainerd/Wav2Vec2-large-xlsr-hindi'
LIVEKIT_TURN_DETECTOR_ID = 'livekit/turn-detector'
SMART_TURN_REPO = 'pipecat-ai/smart-turn-v3'
SMART_TURN_ONNX_FILENAME = 'smart-turn-v3.2-cpu.onnx'
DATASET_REPO = 'pipecat-ai/smart-turn-data-v3.2-train'
DATASET_NUM_SHARDS_TOTAL = 83
EASY_TURN_REPO = 'ASLP-lab/Easy-Turn'
ENCODER_IDS = {'whisper_tiny': WHISPER_TINY_ID, 'whisper_base': WHISPER_BASE_ID, 'wav2vec2': WAV2VEC2_ID}