package com.sherpaonnxofflinestt

/**
 * Centralized configuration constants for STT module.
 * All magic numbers and default values should be defined here.
 *
 * NOTE: This file is mirrored in iOS (ios/STTConstants.swift).
 * Any changes here MUST be mirrored in iOS to maintain platform parity.
 * Both files should have identical values and equivalent constant names.
 */
object STTConstants {
    // =====================
    // Audio Processing
    // =====================

    /** Default sample rate in Hz */
    const val DEFAULT_SAMPLE_RATE = 16000

    /** VAD frame size in samples (10ms at 16kHz) */
    const val VAD_HOP_SIZE = 160

    /** Maximum denoiser chunk size in samples (5 seconds at 16kHz) */
    const val MAX_DENOISER_CHUNK_SAMPLES = 5 * DEFAULT_SAMPLE_RATE

    /** Chunk size for audio recording in milliseconds */
    const val AUDIO_CHUNK_SIZE_MS = 80

    /** Minimum peak amplitude to process STT (avoid noise hallucinations) */
    const val MIN_PEAK_FOR_STT = 0.01f

    /** Target peak amplitude for audio normalization */
    const val TARGET_PEAK_AMPLITUDE = 0.9f

    // =====================
    // VAD Defaults
    // =====================

    /** Default VAD speech probability threshold */
    const val DEFAULT_VAD_THRESHOLD = 0.5f

    /** Minimum speech duration in milliseconds to trigger recognition */
    const val DEFAULT_MIN_SPEECH_MS = 300

    /** Minimum silence duration in milliseconds to end speech segment */
    const val DEFAULT_MIN_SILENCE_MS = 500

    /** Maximum speech duration in milliseconds before forced segment break */
    const val DEFAULT_MAX_SPEECH_MS = 30000

    /** Padding around speech segments in milliseconds */
    const val DEFAULT_SPEECH_PADDING_MS = 100

    /** Threshold adjustment for aggressive VAD mode */
    const val VAD_MODE_AGGRESSIVE_ADJUSTMENT = 0.15f

    /** Threshold adjustment for sensitive VAD mode */
    const val VAD_MODE_SENSITIVE_ADJUSTMENT = -0.15f

    /** Maximum VAD threshold allowed */
    const val VAD_THRESHOLD_MAX = 0.9f

    /** Minimum VAD threshold allowed */
    const val VAD_THRESHOLD_MIN = 0.1f

    // =====================
    // Diarization Defaults
    // =====================

    /** Default speaker similarity threshold for assignment (lenient) */
    const val DEFAULT_SPEAKER_THRESHOLD = 0.55f

    /** Default embedding similarity threshold for profile update (strict) */
    const val DEFAULT_EMBEDDING_THRESHOLD = 0.70f

    /** Maximum number of tracked speakers (LRU eviction) */
    const val DEFAULT_MAX_SPEAKERS = 10

    /** Minimum speech duration for speaker creation in milliseconds */
    const val DEFAULT_MIN_SPEECH_DURATION_MS = 800

    /** Minimum speech duration for embedding update in milliseconds */
    const val DEFAULT_MIN_EMBEDDING_UPDATE_MS = 500

    /** Weight for running average embedding update */
    const val DEFAULT_EMBEDDING_AVERAGE_WEIGHT = 0.2f

    /** Number of onboarding segments for first speaker stabilization */
    const val DEFAULT_ONBOARDING_SEGMENTS = 3

    /** Minimum accumulated speech before creating new speaker in milliseconds */
    const val DEFAULT_MIN_SPEECH_FOR_NEW_SPEAKER = 3000

    /** Diarization window size for sub-segment processing in milliseconds */
    const val DIARIZATION_WINDOW_SIZE_MS = 1500

    /** Diarization window step (50% overlap) in milliseconds */
    const val DIARIZATION_WINDOW_STEP_MS = 750

    // =====================
    // Streaming State
    // =====================

    /** Default confidence threshold for confirming volatile text */
    const val DEFAULT_CONFIRMATION_THRESHOLD = 0.85f

    /** Minimum audio context in milliseconds before confirmation */
    const val DEFAULT_MIN_CONTEXT_MS = 5000L

    // =====================
    // Embedding Validation
    // =====================

    /** Minimum embedding norm for valid embedding */
    const val MIN_EMBEDDING_NORM = 0.001f

    /** Quality normalization factor for embedding variance */
    const val EMBEDDING_QUALITY_FACTOR = 10f

    // =====================
    // Pyannote Defaults
    // =====================

    /** Default clustering threshold for auto speaker detection */
    const val DEFAULT_CLUSTERING_THRESHOLD = 0.5f

    /** Default minimum duration for speech segment in seconds */
    const val DEFAULT_MIN_DURATION_ON = 0.5f

    /** Default minimum duration for silence segment in seconds */
    const val DEFAULT_MIN_DURATION_OFF = 0.3f

    // =====================
    // Streaming Endpoint Detection
    // =====================

    /** Rule 1: Minimum trailing silence in seconds (longer pause detection) */
    const val DEFAULT_RULE1_MIN_TRAILING_SILENCE = 2.4f

    /** Rule 2: Minimum trailing silence in seconds (shorter pause detection) */
    const val DEFAULT_RULE2_MIN_TRAILING_SILENCE = 1.2f

    /** Rule 3: Minimum utterance length in seconds before endpoint */
    const val DEFAULT_RULE3_MIN_UTTERANCE_LENGTH = 20.0f

    // =====================
    // Model Inference
    // =====================

    /** Default number of threads for STT inference */
    const val DEFAULT_STT_NUM_THREADS = 4

    /** Default number of threads for embedding/denoiser inference */
    const val DEFAULT_AUXILIARY_NUM_THREADS = 2

    /** Default ONNX runtime provider */
    const val DEFAULT_PROVIDER = "cpu"
}