Toshimasa Matsuoka commited on
Upload 24k/config.yaml with huggingface_hub
Browse files- 24k/config.yaml +46 -0
24k/config.yaml
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 24 kHz production config (STFT mask path). Matches the trained 24k checkpoint.
|
| 2 |
+
seed: 0
|
| 3 |
+
device: auto
|
| 4 |
+
|
| 5 |
+
stft:
|
| 6 |
+
sample_rate: 24000
|
| 7 |
+
n_fft: 2048
|
| 8 |
+
hop_length: 300
|
| 9 |
+
win_length: 1200
|
| 10 |
+
log_eps: 1.0e-5
|
| 11 |
+
|
| 12 |
+
model:
|
| 13 |
+
base_channels: 32
|
| 14 |
+
depth: 4
|
| 15 |
+
gru_hidden: 256
|
| 16 |
+
gru_reduce: 32
|
| 17 |
+
|
| 18 |
+
norm:
|
| 19 |
+
mean: -3.314941
|
| 20 |
+
std: 2.795372
|
| 21 |
+
|
| 22 |
+
aug:
|
| 23 |
+
snr_db: [0.0, 20.0]
|
| 24 |
+
reverb_prob: 0.5
|
| 25 |
+
noise_prob: 1.0
|
| 26 |
+
remove_reverb: false
|
| 27 |
+
|
| 28 |
+
data:
|
| 29 |
+
clean_dirs: [data/clean_24k] # e.g. LJSpeech + JVS resampled to 24k
|
| 30 |
+
noise_dir: data/noise
|
| 31 |
+
rir_dir: data/rir
|
| 32 |
+
segment_frames: 128
|
| 33 |
+
val_fraction: 0.05
|
| 34 |
+
num_workers: 16
|
| 35 |
+
dummy: false
|
| 36 |
+
dummy_size: 64
|
| 37 |
+
|
| 38 |
+
train:
|
| 39 |
+
batch_size: 16
|
| 40 |
+
lr: 2.0e-4
|
| 41 |
+
max_steps: 200000
|
| 42 |
+
log_every: 100
|
| 43 |
+
val_every: 2000
|
| 44 |
+
ckpt_every: 2000
|
| 45 |
+
ckpt_dir: outputs/ckpt_24k
|
| 46 |
+
amp: true
|