inference.py

import os
import json
import numpy as np
import scipy.io.wavfile as wavfile
from onnxruntime import InferenceSession
from phonemizer import phonemize

# === Step 1: Load phoneme-to-ID vocabulary ===
CONFIG_PATH = "./config_kokoro.json"  # Download this from Hugging Face: Kokoro-82M/config.json
with open(CONFIG_PATH, "r", encoding="utf-8") as f:
    config = json.load(f)
phoneme_to_id = config["vocab"]

# === Step 2: Convert text to phonemes using espeak-ng ===
text = "Hi how are you, what is your name. tell me something"

phonemes = phonemize(
    text,
    language="en-us",
    backend="espeak",
    strip=True,
    preserve_punctuation=True,
    with_stress=True
)

# === Step 3: Filter out unsupported phonemes and convert to token IDs ===
phonemes = "".join(p for p in phonemes if p in phoneme_to_id)
print("Phonemes:", phonemes)

tokens = [phoneme_to_id[p] for p in phonemes]
print("Token IDs:", tokens)

# === Step 4: Prepare style embedding and input IDs ===
assert len(tokens) <= 510, "Token sequence too long (max 510 phonemes)"

voices = np.fromfile('./voices/af.bin', dtype=np.float32).reshape(-1, 1, 256)
ref_s = voices[len(tokens)]  # Select style vector based on token length

tokens = [[0, *tokens, 0]]  # Add padding tokens at the beginning and end

# === Step 5: Run ONNX model inference ===
model_name = 'model.onnx'
sess = InferenceSession(os.path.join('onnx', model_name))

audio = sess.run(None, {
    'input_ids': tokens,
    'style': ref_s,
    'speed': np.ones(1, dtype=np.float32),
})[0]

# === Step 6: Save output audio as a 24kHz WAV file ===
wavfile.write('audio.wav', 24000, audio[0])
print("✅ Audio saved to audio.wav")