config.example.yaml

source_language: ru
target_language: sr-Latn
work_directory: work
output_directory: output

tools:
  ffmpeg: tools/ffmpeg/ffmpeg.exe
  ffprobe: tools/ffmpeg/ffprobe.exe
  yt_dlp: tools/yt-dlp/yt-dlp.exe

audio:
  asr_sample_rate: 16000
  mix_sample_rate: 48000

asr:
  # Parakeet won the real Russian-film A/B on speed and hallucination resistance
  # at effectively tied CER/WER. whisper.cpp remains a selectable CUDA fallback.
  engine: parakeet
  executable: tools/whisper/whisper-cli.exe
  model: models/ggml-large-v3.bin
  # Disabling temperature fallback avoids a common source of confident stock-text
  # hallucinations. Physical VAD and the quality gate remain active as a second layer.
  extra_args: "--threads 8 --no-fallback"
  # Project-local primary backend. setup-tools.bat installs both the native CLI
  # and the pinned model; setup-asr-parakeet.bat can verify/reinstall it alone.
  parakeet_executable: tools/whisper/parakeet-cli.exe
  parakeet_model: models/ggml-parakeet-tdt-0.6b-v3-q8_0.bin
  parakeet_extra_args: "--threads 8"
  parakeet_segment_max_seconds: 8
  tiny_diarize: false
  vad: true
  # Physical Silero windows keep wall-clock timestamps stable. embedded_vad is
  # retained only as a legacy diagnostic mode because whisper.cpp can stretch
  # embedded-VAD segments across removed silence.
  vad_executable: tools/whisper/whisper-vad-speech-segments.exe
  embedded_vad: false
  vad_model: models/ggml-silero-v6.2.0.bin
  vad_threshold: 0.5
  vad_min_silence_ms: 250
  vad_max_speech_seconds: 30
  # Gaps below this threshold stay in one linguistic phrase; longer pauses become
  # hard scheduling boundaries which translation/TTS are not allowed to erase.
  vad_phrase_silence_ms: 250
  vad_speech_padding_ms: 80
  # Whisper can stretch a segment timestamp across a long silent interval.
  # Refine only boundaries that fall inside measured silence in the ASR track.
  refine_timestamps: true
  timestamp_silence_threshold_db: -35
  timestamp_min_silence_ms: 250
  timestamp_speech_padding_ms: 50
  chunk_minutes: 15
  chunk_overlap_seconds: 3
  retry_chunk_minutes: 3
  max_context_tokens: 0
  max_pattern_repeats: 8
  max_segments_per_minute: 45

separation:
  mode: demucs
  engine: demucs
  executable: tools/separation/.venv/Scripts/python.exe
  model: htdemucs_ft
  device: cuda
  # Softer, reconstruction-safe separation. Higher shifts/overlap improve stability
  # at the cost of a slower separation stage.
  other_method: minus
  shifts: 2
  overlap: 0.50
  arguments: '-m demucs --two-stems vocals -n "{model}" --device "{device}" -o "{output_dir}" "{input}"'
  timeout_hours: 12

prepare:
  merge_below_seconds: 2.2
  max_merged_seconds: 10.0
  max_gap_seconds: 0.65
  preserve_pause_seconds: 0.25

source_subtitles:
  maximum_download_mb: 16
  match_padding_seconds: 1.0
  local_search_seconds: 8.0
  maximum_offset_seconds: 120
  minimum_time_scale: 0.95
  maximum_time_scale: 1.05
  minimum_alignment_score: 0.15
  # Keep real ASR-only material (for example songs omitted by the SRT).
  # Duplicate/corrupt ASR fragments inside a matched SRT cue are still replaced.
  preserve_unmatched_asr: true

translation:
  # ollama, codex-cli, or claude-cli. CLI engines use the authenticated
  # desktop subscription by default and do not use API-key environment variables.
  engine: codex-cli
  # dub-video.bat first reuses a running system Ollama on 127.0.0.1:11434.
  # This isolated project-local endpoint remains the fallback when none is running.
  endpoint: http://127.0.0.1:11435
  model: translategemma:12b
  # Leave these empty to use "codex"/"claude" from PATH and the CLI's default model.
  cli_executable:
  cli_model:
  cli_extra_args:
  cli_timeout_minutes: 20
  cli_subscription_only: true
  batch_size: 6
  cli_batch_size: 80
  cli_context_segments: 20
  temperature: 0.0
  context_length: 8192
  max_output_tokens: 2048
  think: false
  prompt: prompts/translation-system.txt
  shorten_prompt: prompts/shorten-system.txt
  # Optional <target_language>.txt rules are appended to translate, verify and shorten prompts.
  target_prompt_directory: prompts/targets
  # A second model call is optional. Deterministic script/shape checks and retries
  # still run when this is false; enable it only when the selected model benefits.
  verify: false
  verification_prompt: prompts/verify-translation-system.txt
  max_retries: 2

# Local Serbian-capable AR + diffusion TTS. The persistent worker loads the
# 4-bit model once per stage and uses about 8 GB VRAM. Single mode creates one
# stable neutral profile when voice is empty. Speakers mode uses consented
# voices/SPEAKER_XX.wav references or stable generated per-speaker profiles.
tts:
  engine: kugelaudio
  voice:
  voice_mode: single
  speaker_voice_directory: voices
  seed: 20260712
  executable: tools/tts/kugelaudio/.venv/Scripts/python.exe
  # Deferring the first end token by one acoustic frame lets KugelAudio render
  # the final phoneme release and then stop naturally instead of cutting at an
  # active 133 ms frame boundary.
  arguments: 'scripts/kugelaudio-worker.py --model kugelaudio/kugelaudio-0-open --cache tools/tts/kugelaudio/cache --4bit --cfg-scale 3.5 --end-tail-frames 1'
  timeout_minutes: 30

timing:
  min_tempo: 0.88
  max_tempo: 1.22
  hard_max_tempo: 1.60
  short_phrase_seconds: 2.0
  short_phrase_hard_max_tempo: 2.0
  max_overflow_seconds: 0.6
  max_shorten_attempts: 3
  trim_tts_silence: true
  silence_threshold_db: -45
  silence_padding_ms: 50
  # Guaranteed post-roll before tempo fitting; prevents clipped final phonemes.
  tail_padding_ms: 160

mix:
  original_volume: 1.0
  duck_volume: 0.06
  voice_volume: 1.0
  original_voice_volume: 0.10
  attack_ms: 150
  release_ms: 300
  source_guard_ms: 250

video:
  audio_bitrate: 192k
  overwrite: true