Configuration.py

from Imports import *

SAMPLE_RATE      = 16000
FRAME_STEP       = 256
FFT_LENGTH       = 1024
NUM_MEL_BINS     = 80
LOWER_EDGE_HERTZ = 80.0
UPPER_EDGE_HERTZ = 7600.0
R           = 3      # reduction factor — predict R mel frames per decoder step
PRE_EMPH    = 0.97   # pre-emphasis coefficient
GRAD_CLIP   = 5.0    # gradient clip norm

# ── NEW: static shapes for training ───────────────────────────────────────────
MAX_TEXT_LEN = 200          # pad all text sequences to this
MAX_MEL_LEN  = 600          # pad all mel sequences to this
MAX_WAV_LEN  = MAX_MEL_LEN * FRAME_STEP   # = 222720 samples

VOCAB        = list("abcdefghijklmnopqrstuvwxyz .,!?-'\"")
PAD_TOKEN    = '<PAD>'
EOS_TOKEN    = '<EOS>'
vocab_list   = [PAD_TOKEN, EOS_TOKEN] + VOCAB
char2id      = {c: i for i, c in enumerate(vocab_list)}
id2char      = {i: c for c, i in char2id.items()}
VOCAB_SIZE   = len(char2id)
PAD_ID       = char2id[PAD_TOKEN]
EOS_ID       = char2id[EOS_TOKEN]

keys_tensor   = tf.constant(list(char2id.keys()))
values_tensor = tf.constant(list(char2id.values()), dtype=tf.int32)
char_to_id_table = tf.lookup.StaticHashTable(
    tf.lookup.KeyValueTensorInitializer(keys_tensor, values_tensor),
    default_value=PAD_ID
)
keys_tensor2   = tf.constant(list(id2char.keys()),   dtype=tf.int32)
values_tensor2 = tf.constant(list(id2char.values()))
id_to_char_table = tf.lookup.StaticHashTable(
    tf.lookup.KeyValueTensorInitializer(keys_tensor2, values_tensor2),
    default_value='?'
)
print(VOCAB_SIZE)