armenian-video-dubbing

Configuration Reference

Overview

All pipeline settings are centralized in configs/config.yaml. This document describes every configuration option.

Project

project:
  name: "armenian-video-dubbing"
  version: "0.1.0"
  root: "."
  device: "cuda"    # cuda, cpu, mps
  dtype: "float16"  # float16, bfloat16, float32
  seed: 42

Key	Type	Description
device	string	Compute device. cuda (NVIDIA GPU), cpu, or mps (Apple Silicon)
dtype	string	Default tensor precision
seed	int	Random seed for reproducibility

ASR

asr:
  backend: "whisper"    # whisper or nemo
  whisper:
    model: "large-v3"
    language: "hy"
    beam_size: 5
    vad_filter: true
    word_timestamps: true
    batch_size: 16

Key	Type	Description
backend	string	ASR engine: whisper or nemo
model	string	Whisper model size
language	string	Target language code
beam_size	int	Beam search width
vad_filter	bool	Voice Activity Detection filtering
word_timestamps	bool	Enable word-level timestamps
batch_size	int	Inference batch size

Translation

translation:
  model: "facebook/seamless-m4t-v2-large"
  source_lang: "eng"
  target_lang: "hye"
  dialect: "eastern"    # eastern or western
  max_length: 512
  num_beams: 5

Key	Type	Description
dialect	string	eastern (hye) or western (hyw)
source_lang	string	ISO 639-3 source language
target_lang	string	ISO 639-3 target language
max_length	int	Maximum sequence length
num_beams	int	Beam search width

TTS

tts:
  backend: "fish-speech"    # fish-speech or cosyvoice
  fish_speech:
    sample_rate: 44100
    max_new_tokens: 2048
    top_p: 0.7
    temperature: 0.7
    repetition_penalty: 1.2
    emotion_tags: true
    reference_audio_seconds: 10
    chunk_length: 200

Key	Type	Description
backend	string	TTS engine
emotion_tags	bool	Enable emotion-aware synthesis
reference_audio_seconds	int	Duration of voice reference clip
chunk_length	int	Characters per synthesis chunk
temperature	float	Sampling temperature (creativity)
top_p	float	Nucleus sampling threshold

Lip-Sync

lipsync:
  model: "musetalk"
  fps: 25
  face_det_batch_size: 8
  bbox_shift: 0
  batch_size: 16
  use_float16: true
  face_enhancement: true

Key	Type	Description
model	string	Lip-sync model (musetalk)
fps	int	Video frame rate
bbox_shift	int	Vertical face bounding box adjustment
face_enhancement	bool	Enable CodeFormer post-processing

Audio Post-Processing

audio:
  demucs:
    model: "htdemucs_ft"
    segment: 7.8
    overlap: 0.25
    shifts: 1
  sample_rate: 44100
  loudness_target: -14.0    # LUFS
  crossfade_ms: 50

Key	Type	Description
demucs.model	string	Demucs variant for source separation
loudness_target	float	Target loudness in LUFS
crossfade_ms	int	Crossfade duration between segments

Duration Matching

timing:
  max_stretch_ratio: 1.25
  min_compress_ratio: 0.80
  silence_threshold_db: -40
  min_pause_ms: 150
  method: "rubberband"

Key	Type	Description
max_stretch_ratio	float	Maximum stretching factor
min_compress_ratio	float	Maximum compression factor
method	string	Time-stretch algorithm (rubberband or wsola)

Inference

inference:
  batch_video_max_concurrent: 2
  gpu_memory_fraction: 0.9
  enable_quantization: true
  quantization_bits: 4

Key	Type	Description
enable_quantization	bool	Enable model quantization
quantization_bits	int	Quantization precision (4 or 8)
gpu_memory_fraction	float	Maximum GPU memory usage

Ethics

ethics:
  add_watermark: true
  watermark_text: "AI-Dubbed"
  watermark_opacity: 0.3
  consent_required: true
  consent_log_path: "logs/voice_consent.json"

Key	Type	Description
add_watermark	bool	Overlay AI watermark on output
consent_required	bool	Require voice consent logging
consent_log_path	string	Path to consent log file

Environment Variables

Settings can also be controlled via environment variables (see .env.example):

Variable	Description
HF_TOKEN	HuggingFace access token
ARMTTS_API_KEY	API authentication key
ARMTTS_CONFIG	Path to config file
CUDA_VISIBLE_DEVICES	GPU device selection
WANDB_API_KEY	Weights & Biases tracking

This site is open source. Improve this page.