config.yaml

exp_name: shona_f5_tts-m150-n150-full-lr1em5-ep30-April-26-2026_02+51PM-b60c1d4
output_dir: /root/project/runs/shona_f5_tts/shona_f5_tts-m150-n150-full-lr1em5-ep30-April-26-2026_02+51PM-b60c1d4/checkpoints
seed: 42
model:
  type: F5TTS_v1_Base
  pretrained: SWivid/F5-TTS
  pretrained_ckpt: F5TTS_v1_Base/model_1250000.safetensors
  vocab_path: /root/project/datasets/sna-manasseh-150-raw/f5_prepared/default/vocab.txt
  use_lora: false
  dim: 1024
  depth: 22
  heads: 16
  ff_mult: 2
  text_dim: 512
  conv_layers: 4
dataset:
  name: shona_f5_tts-m150-n150-full-lr1em5-ep30-April-26-2026_02+51PM-b60c1d4
  dir: /root/project/datasets/sna-manasseh-150-raw/f5_prepared/default
  metadata: metadata.csv
  sample_rate: 24000
training:
  epochs: 30
  learning_rate: 1.0e-05
  weight_decay: 0.01
  max_grad_norm: 1.0
  grad_accumulation_steps: 1
  batch_size_per_gpu: 1600
  batch_size_type: frame
  max_samples: 64
  num_warmup_updates: 500
  lr_scheduler: cosine
  use_ema: true
  ema_decay: 0.9999
  mixed_precision: fp16
  num_workers: 4
checkpoint:
  save_per_updates: 66
  last_per_steps: 10000
  keep_last_n: 5
logging:
  type: wandb
  log_per_updates: 50
  wandb_project: f5-tts-shona-finetuning
  wandb_run_name: shona_f5_tts-m150-n150-full-lr1em5-ep30-April-26-2026_02+51PM-b60c1d4