config.yaml

# Auto Subtitle Generator Configuration
# -------------------------------------
# This configuration file controls the behavior of the AI Audio Subtitle Generator pipeline.
# Include settings for Whisper AI transcription, hallucination filtering, and NLLB translation.

# -----------------------------------------------------------------------------
# Whisper AI Settings
# -----------------------------------------------------------------------------
# Configuration for the OpenAI Whisper automatic speech recognition model.
whisper:
  # Model Size: Determines accuracy and resource usage.
  # Available Options:
  # - "tiny" / "tiny.en"       (Fastest, lowest accuracy, ~1GB VRAM)
  # - "base" / "base.en"       (Balanced for simple audio, ~1-2GB VRAM)
  # - "small" / "small.en"     (Good general purpose, ~2-3GB VRAM)
  # - "medium" / "medium.en"   (High accuracy, ~5GB VRAM)
  # - "large-v3"               (Best accuracy, requires ~10GB VRAM)
  model_size: "large-v3"
  
  # Forced Language: Force Whisper to assume this language (e.g., "ro", "en").
  # Leave null to use auto-detection.
  language: null
  
  # Use Prompt: Boolean validation to enable contextual prompting.
  # Set to true to provide Whisper with context about the audio (e.g., specific terminology).
  use_prompt: true
  
  # Custom Prompt: Text string to guide the model's style or vocabulary.
  # Leave empty to use the system default optimization prompt.
  custom_prompt: "This video contains speech in multiple languages including Romanian, English, French, Italian, German, and Spanish."
  
  # Custom Prompt Priority: If true, the custom_prompt above overrides everything.
  # If false, the script will still use the "Smart Auto-Detect" bias (e.g. adding 'Primarily in RO').
  custom_prompt_priority: false
 
  # Force Detected Language: If true, the result of the "Robust Scan" is forced as the primary language.
  # This prevents Whisper from locking onto an English intro.
  force_detected_language: false
 
  # Use Vocal Separation: If true, use BS-Roformer to clean the audio.
  # Set to false if the isolated track sounds 'metallic' or missing content.
  use_vocal_separation: true
 
# -----------------------------------------------------------------------------
# Hallucination & Noise Filters
# -----------------------------------------------------------------------------
# Filters to post-process Whisper output and remove common AI artifacts, 
# such as repetitive text generated during silence or background music.
hallucinations:
  # Silence Threshold: Probability threshold (0.0 - 1.0) to discard low-confidence segments.
  # Set to 0.85 for old/noisy recordings to prevent valid speech from being skipped.
  silence_threshold: 0.85
  
  # Repetition Threshold: non-negative integer.
  # Flags a segment as invalid if the exact same text appears consecutively this many times.
  repetition_threshold: 15
  
  # Known Phrases: List of specific strings to continually ignore.
  # Useful for removing recurring "subscribe" channel calls or AI hallucinations.
  known_phrases:
    - "nu uitați să dați like"
    - "nu uitati sa dati like"
    - "să lăsați un comentariu"
    - "sa lasati un comentariu"
    - "să distribuiți"
    - "sa distribuiti"
    - "abonați-vă la canal"
    - "abonati-va la canal"
    - "mulțumesc pentru vizionare"
    - "multumesc pentru vizionare"
    - "thank you for watching"
    - "thanks for watching"
    - "don't forget to subscribe"
    - "please subscribe"
    - "like and subscribe"
    - "hit the like button"
    - "leave a comment"
    - "share this video"
    - "see you in the next"
    - "bye bye"
    - "merci d'avoir regardé"
    - "n'oubliez pas de vous abonner"
    - "laissez un commentaire"
    - "à bientôt"
    - "danke fürs zuschauen"
    - "vergisst nicht zu abonnieren"
    - "gracias por ver"
    - "no olvides suscribirte"
    - "grazie per aver guardato"
    - "non dimenticare di iscriverti"
 
# -----------------------------------------------------------------------------
# NLLB Translation Targets
# -----------------------------------------------------------------------------
# Defines the target languages for the NLLB (No Language Left Behind) translator.
# The keys (e.g., 'en', 'fr') are used in output filenames.
# Values map to the specific NLLB model code and a human-readable label.
target_languages:
  # Major Global Languages
  en: {code: "eng_Latn", label: "English"}
  zh: {code: "zho_Hans", label: "Chinese (Simplified)"}
  es: {code: "spa_Latn", label: "Spanish"}
  fr: {code: "fra_Latn", label: "French"}
  ar: {code: "arb_Arab", label: "Arabic"}
  ru: {code: "rus_Cyrl", label: "Russian"}
  pt: {code: "por_Latn", label: "Portuguese"}
  de: {code: "deu_Latn", label: "German"}
  ja: {code: "jpn_Jpan", label: "Japanese"}
  hi: {code: "hin_Deva", label: "Hindi"}
 
  # European & Regional
  ro: {code: "ron_Latn", label: "Romanian"}
  tr: {code: "tur_Latn", label: "Turkish"}
  it: {code: "ita_Latn", label: "Italian"}
  ko: {code: "kor_Hang", label: "Korean"}
  vi: {code: "vie_Latn", label: "Vietnamese"}
  pl: {code: "pol_Latn", label: "Polish"}
  nl: {code: "nld_Latn", label: "Dutch"}
  id: {code: "ind_Latn", label: "Indonesian"}
  uk: {code: "ukr_Cyrl", label: "Ukrainian"}
  th: {code: "tha_Thai", label: "Thai"}
  cs: {code: "ces_Latn", label: "Czech"}
  hu: {code: "hun_Latn", label: "Hungarian"}
  sv: {code: "swe_Latn", label: "Swedish"}
  el: {code: "ell_Grek", label: "Greek"}
  
  # Nordic & Others
  da: {code: "dan_Latn", label: "Danish"}
  fi: {code: "fin_Latn", label: "Finnish"}
  no: {code: "nob_Latn", label: "Norwegian"}
  bg: {code: "bul_Cyrl", label: "Bulgarian"}
  hr: {code: "hrv_Latn", label: "Croatian"}
  sr: {code: "srp_Cyrl", label: "Serbian"}
  sk: {code: "slk_Latn", label: "Slovak"}
  sl: {code: "slv_Latn", label: "Slovenian"}
  lt: {code: "lit_Latn", label: "Lithuanian"}
  lv: {code: "lvs_Latn", label: "Latvian"}
  et: {code: "est_Latn", label: "Estonian"}
  he: {code: "heb_Hebr", label: "Hebrew"}
  fa: {code: "pes_Arab", label: "Persian"}
 
# -----------------------------------------------------------------------------
# NLLB Translation Quality Settings
# -----------------------------------------------------------------------------
# Advanced parameters for the NLLB-200 translation model.
# These control the trade-off between speed and linguistic excellence.
nllb:
  # Number of Beams: Increases hypothesis exploration.
  # Available Options:
  # - 5  (Standard high-quality)
  # - 10 (Ultra-Quality: Best linguistic flow, slower)
  num_beams: 10
 
  # Length Penalty: Encourages the model to produce longer/more descriptive translations.
  # - 1.0 (Neutral)
  # - 1.1 (Slightly more descriptive - Recommended for high quality)
  length_penalty: 1.1
 
  # No Repeat N-Gram Size: Prevents the model from repeating the same word chains.
  # - 0 (Off - Let the model decide)
  # - 4 (Gentle repetition prevention for long narratives)
  no_repeat_ngram_size: 0
 
  # Repetition Penalty: Penalizes words already generated.
  # - 1.0 (Neutral - Recommended for 3.3B model)
  repetition_penalty: 1.0
 
# -----------------------------------------------------------------------------
# File Type Settings
# -----------------------------------------------------------------------------
# Supported video extensions to scan for.
file_types:
  extensions:
    - ".mp4"
    - ".mkv"
    - ".mov"
    - ".avi"
    - ".webm"
    - ".flv"
    - ".m4v"
    - ".ts"
    - ".mts"
 
# -----------------------------------------------------------------------------
# Model Configuration
# -----------------------------------------------------------------------------
# Specific model IDs or paths for the AI components.
models:
  # NLLB Translation Model (HuggingFace ID)
  # Available Options:
  # - "facebook/nllb-200-3.3B"            (High quality, slower, ~8-10GB VRAM)
  # - "facebook/nllb-200-1.3B"            (Balanced, faster, ~4-5GB VRAM)
  # - "facebook/nllb-200-distilled-600M"  (Fastest, lower quality, ~1-2GB VRAM)
  nllb: "facebook/nllb-200-3.3B"
  
  # Audio Separator Model (BS-Roformer Checkpoint)
  # Common options supported by python-audio-separator:
  # - "model_bs_roformer_ep_317_sdr_12.9755.ckpt" (Best vocal isolation)
  # - "UVR-MDX-Net-Inst_HQ_3.onnx"                (Good general purpose)
  # Ensure the model filename matches what is expected by the library.
  audio_separator: "model_bs_roformer_ep_317_sdr_12.9755.ckpt"
 
# -----------------------------------------------------------------------------
# Performance Tuning
# -----------------------------------------------------------------------------
# Manual overrides for system optimization. 
# Leave these commented out or set to null to let the auto-optimizer decide based on hardware.
performance:
  # Target VRAM in GB for dynamic scaling (Recommended: 28 for 32GB cards)
  max_vram_usage_gb: null
 
  # Whisper workers (Parallel transcription: 10+ recommended for 9950X)
  whisper_workers: null
  
  # Beam size for Whisper (higher = more accurate, slower)
  whisper_beam: null
  
  # Batch size for NLLB translation (will be dynamically overridden if null)
  nllb_batch: null
  
  # FFmpeg threads (0 = auto)
  ffmpeg_threads: null
 
# -----------------------------------------------------------------------------
# Voice Activity Detection (VAD)
# -----------------------------------------------------------------------------
# Parameters for Whisper's internal VAD filter to detecting speech vs silence.
vad:
  # Minimum duration of silence (in ms) to consider a break in speech.
  min_silence_duration_ms: 500
 
# ------------------------------------------------------------------------------
# DEBUGGING
# ------------------------------------------------------------------------------
# debug_logging:
#   - false: Standard output (Recommended).
#   - true:  Detailed terminal logs for troubleshooting.
debug_logging: false