# configs/indic_parler_bhojpuri.yaml
#
# Plan B — Language Adaptation: Teach Bhojpuri phoneme patterns to Indic Parler TTS.
#
# Goal: After training, the model synthesizes Bhojpuri when given Bhojpuri text.
#       Description strings still control prosody / voice quality.
#       No specific speaker identity is baked in.
#
# Usage:
#   python scripts/prepare_bhojpuri_dataset.py --mode language --max-hours 10
#   python scripts/preprocess_bhojpuri_dac.py
#   python scripts/train.py --config configs/indic_parler_bhojpuri.yaml

model:
  model_type: indic_parler_tts
  hf_repo: ai4bharat/indic-parler-tts

# No fixed speaker_description — varied per-clip descriptions from JSONL.
# This teaches phoneme patterns independent of voice style.

lora:
  rank: 8
  alpha: 16                        # scale = 2.0 — Bhojpuri needs larger updates than speaker adapt
  target_modules:
    # Self-attention drives phoneme sequence learning
    - decoder.layers.*.self_attn.q
    - decoder.layers.*.self_attn.k
    - decoder.layers.*.self_attn.v
    - decoder.layers.*.self_attn.out
    # FFN layers critical for new phoneme distributions
    - decoder.layers.*.fc1
    - decoder.layers.*.fc2
    # Cross-attention so description conditioning still works
    - decoder.layers.*.cross_attn.q
    - decoder.layers.*.cross_attn.v

frozen:
  - text_encoder                   # T5 fully frozen — handles any description already
  - embed_prompts                  # prompt tokenizer embedding frozen
  - dac                            # DAC decoder frozen

data:
  processor: indic_parler
  train_jsonl: /Users/akashsingh/Documents/exps/bhojpuri/train.jsonl
  val_jsonl:   /Users/akashsingh/Documents/exps/bhojpuri/val.jsonl
  max_audio_length_s: 12.0
  num_codebooks: 9
  max_frames: 1032                 # 12s × 86Hz = 1032 frames

trainer:
  loss_fn: parler_tts_loss
  run_name: parler-bhojpuri-language
  learning_rate: 1.0e-5            # higher than speaker adapt — new language needs more movement
  num_epochs: 3
  batch_size: 1
  grad_accumulation: 8             # effective batch = 8
  label_smoothing: 0.1             # diversity across Bhojpuri phonemes
  grad_clip: 1.0
  warmup_steps: 100
  eval_every_n_steps: 200
  save_every_n_steps: 500
  output_dir: /Users/akashsingh/Documents/exps/checkpoints/parler-bhojpuri-language
  tensorboard_dir: /Users/akashsingh/Documents/exps/tensorboard/parler-bhojpuri-language
  log_file: /Users/akashsingh/Documents/exps/checkpoints/parler-bhojpuri-language/train_log.jsonl

# Bhojpuri-specific stopping criterion (manual eval every 200 steps):
#   1. Run inference on Bhojpuri test sentence — listen for correct phonemes
#   2. Run inference on Hindi/Tamil/Telugu sentences — must remain intelligible
#   3. Val loss on Bhojpuri held-out set must decrease
#
# Risk: Hindi accent bleed (Devanagari overlap).
# Mitigation: Mix 200-300 Hindi samples into training data (set in prepare script).

eval:
  test_sentences:
    bhojpuri:
      - text: "रउरा के राम राम। आज हम एही गाँव में रहीला।"
        description: "A female speaker delivers speech at a moderate pace. The recording is of very high quality."
      - text: "ई बहुत नीमन बा। हम कल जाइब।"
        description: "A female speaker delivers speech at a moderate pace. The audio quality is excellent."
    hindi_guard:
      - text: "नमस्ते, आप कैसे हैं?"
        description: "Divya's voice is slightly expressive and very animated. She speaks at a moderate pace."
    tamil_guard:
      - text: "வணக்கம், எப்படி இருக்கீங்க?"
        description: "A female speaker delivers speech at a moderate pace. The recording is very clean."

# TensorBoard audio: generates these sentences after each eval step and logs waveforms.
# Launch TensorBoard: tensorboard --logdir /Users/akashsingh/Documents/exps/tensorboard
eval_audio:
  enabled: true
  max_samples: 3
  max_audio_length_s: 8.0
  temperature: 1.0
  test_sentences:
    - text: "रउरा के राम राम। आज हम एही गाँव में रहीला।"
      description: "A female speaker delivers speech at a moderate pace. The recording is of very high quality."
      tag: "bhojpuri_0"
    - text: "ई बहुत नीमन बा। हम कल जाइब।"
      description: "A female speaker delivers speech at a moderate pace. The audio quality is excellent."
      tag: "bhojpuri_1"
    - text: "नमस्ते, आप कैसे हैं?"
      description: "Divya's voice is slightly expressive and very animated. She speaks at a moderate pace."
      tag: "hindi_guard"